YT YouTube2Text

YouTube2Text → Transcripts → AI Agents: The Most Valuable Skill You Can Learn in 2026 (Full Course)

AI Agents: The Most Valuable Skill You Can Learn in 2026 (Full Course) — Transcript

by Open Residency · 32,541 words · 4,932 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Every company needs to put AI at the
0:02forefront of all their operations,
0:03[music] or they're going to lose to a
0:04company that does.
0:05>> That's Remi. He's a young, non-technical
0:07founder [music]
0:08who's a master at explaining AI in a way
0:11you can actually understand [music] and
0:12implement.
0:13>> In order to make the most out of these
0:14agents and make them actually useful,
0:16you need to treat setting [music] them
0:17up like onboarding a real employee. And
0:19you do this through context, tools, and
0:22skills.
0:22>> He shares his screen and shows us
0:24exactly how he [music] builds AI
0:25systems, timeless principles you can
0:27steal, and the easiest way to get
0:29started today.
0:30>> If you're starting from zero, this is
0:31all you need. Just this episode, and
0:33[music] I think you're set.
0:44>> Guys, in this episode, Remi is going to
0:46demystify everything that's going on
0:48with AI, the terms, the tools, the
0:49models, everything. An A-to-Z guide to
0:53know what's going on today in AI. He has
0:55spent the last 2 years living in this
0:57stuff, and he is about to embed in your
0:59brain all things AI that he has learned
1:01in the next couple hours. He's not some
1:02technical engineer either. For the last
1:04couple years, he's been running an ad
1:05agency. This is all self-taught, and now
1:08he's running the AI agent team for one
1:09of the fastest-growing e-com brands in
1:11the world today. And I know this AI
1:13stuff is very overwhelming, so I want
1:15you all to take a deep breath, and by
1:17the end of this episode, you're going to
1:19know exactly what the next steps are.
1:21Let's get into it. So,
1:22>> Oh, yeah.
1:23>> I guess let's just start with, you know,
1:24what are the kind of the two stages of
1:27using AI?
1:28>> So, stage one was chat, right? You'd be
1:33pretty hard-pressed to find anyone,
1:34especially founders and business owners,
1:36that aren't using chat models. So, think
1:38like ChatGPT or Claude Chat or Gemini.
1:42But, stage two is now agents. So, we've
1:45moved from chat to agents.
1:47And the founders and employees that are
1:50now using agents are like nowhere to be
1:51found. They're like 5 to 10 times more
1:53productive in their day.
1:54>> Probably more. Probably more than 5 to
1:5510, to be honest with you.
1:56>> Probably. I mean, even if you take like
1:58the the middle mark, which is like seven
2:00times more productive. Think about that.
2:01That's a week in a working day. And like
2:04five to seven weeks in a week. And then
2:06if you stack that up over like a month,
2:08a year, like it's toast. So, you're
2:10starting to just get this massive divide
2:12opening at the moment with the people
2:15who are using agents and the people who
2:16are stuck using chat.
2:17And you want to be on the agent side.
2:21>> Yeah, I mean, it's unfortunate for some
2:23of the people listening out here that
2:24may or may not be employees, but like a
2:26big thing we talk about is just OPEX,
2:28your actual labor as a percentage of
2:29revenue.
2:30>> Yeah.
2:30>> I think everybody is racing to just get
2:32this down, the percentage of revenue,
2:34and I feel like people are doing it with
2:37just chat, and they don't understand the
2:39true multiplier
2:40>> Yeah.
2:41>> in actually with the agents.
2:42>> And even like you're saying about stuff
2:44for employees, I actually
2:46think that there's a great opportunity
2:48for employees right now, which we'll
2:49touch on later, but I think there's a
2:50there's an amazing opportunity for
2:51employees that take the initiative to
2:53learn this stuff. You end up becoming
2:55like a 100x employee, and it's like you
2:57can now
2:58like one like go-to-market strategist
3:00can now take on like the work of a
3:0210-person team and get paid for probably
3:04about five five-person team now.
3:06>> Yeah, we're going to dive deep into what
3:08I think and what you think the future of
3:10a team looks like, what are the
3:11necessary skills, what does that brain
3:13need to look like, and like I mean,
3:15dude, we're seeing like I think there
3:16there was a company that was a billion
3:17dollars that was one person. Like I
3:18truly believe there's going to be one to
3:21three-person teams doing billions and
3:22billions.
3:23>> I agree. Because there's so much of your
3:25time and day that's like taken up with
3:26just like stupid manual tasks. Like I'll
3:30give you an example. Yesterday, there
3:31was someone from my team was asking for
3:32like some passwords and API keys. And
3:34it's like before I would have had to go
3:35into one password, like find them,
3:38create a share link, share it securely,
3:40and like send a Slack message. So,
3:42instead, I just got my agent to just
3:43like go into one password, like create
3:45the share link, send them on Slack. And
3:47it's like
3:48that would have been maybe like a
3:4910-minute task, realistically. But it's
3:52like
3:53just 30 seconds. And it's like
3:55it just allows you as a founder to spend
3:57your time on the work that's fulfilling
3:58for you and creative and strategic. The
4:00work that AI like can't do. Like the
4:02visionary stuff.
4:03>> Yeah, for sure. Let's dive into, you
4:05know, what does that agent look like?
4:06Let's really, really deep dive into it.
4:08Anybody out here out here that's
4:09listening, I highly recommend that you
4:10either watch this on YouTube or Spotify
4:13so you actually can see the visuals, but
4:15let's break down, you know, agent a bit
4:17more and like what that kind of loop
4:18looks like.
4:20>> 100%. So, the word agent gets thrown
4:22around so much. It's kind of like lost a
4:24lot of meaning now. You just see like if
4:25you open Instagram or X, it's just "I
4:28built an agent that does this, agents
4:29this, agents that."
4:31But, the easiest way to like actually
4:32think about like what an agent is
4:34is I like to think chat is just question
4:37to answer
4:39and an agent is just goal to result. So,
4:41question to answer, goal to result. It's
4:43just the simple way to think about it in
4:44your head.
4:45So,
4:46for example, with a chat, you might ask
4:48it a question, it'll reply, and then you
4:50have to go and do the work. Whereas,
4:53agents, you're giving it a goal, it's
4:55breaking it down into a plan and
4:56executing and then handing you like a
4:57finished product.
4:59And the word agent it actually just
5:01comes from the word agency, which means
5:03like ability to act and do things.
5:05And people think like agent as in like
5:08real estate agent or like travel agent
5:10as like a person, like a guy that's like
5:12a sci-fi type guy that's like living his
5:15own life and then doing work for you.
5:17But, in reality, it's just like comes
5:18from that word agency, which was like
5:20agentic AI, which meant like AI that can
5:23do things for you. And then people just
5:24like now call them agents. The way I
5:27think of it is like chat model is like a
5:29really brilliant consultant. Like you
5:30can go ask it advice, and it would like
5:32give you give you the answer, but you
5:34have to actually go and take that answer
5:36and do the work. Whereas, an agent's
5:37like hiring that brilliant person
5:40in-house that will just then go and do
5:41it for you, too.
5:42>> Yeah. End to end. Like you What do you
5:45want it What does that result look like?
5:47They figure out kind of the work back
5:48schedule and then literally like when
5:50you plus press go and we're going to
5:51talk about underneath the hood like what
5:53it looks like to obviously the piping to
5:54make that work. You just see it just
5:56typing and working and it just gets the
5:58end result.
5:58>> Yeah, cuz I would argue that chat models
6:01questions to answer like can't actually
6:03even take any meaningful work off your
6:05plate. Like maybe a research task, but
6:07realistically like chat models can't
6:08actually take work off your plate.
6:10>> Yeah, let's talk about like a couple of
6:11those things. Let's like kind of
6:13categorize like stuff that maybe does
6:14make sense in chat. Like for me it'll be
6:17just like, you know, I'm staying at this
6:19hotel. I'm going to this studio and then
6:21I'm going to this, you know, should I go
6:22should I book this hotel or this hotel
6:25and I'll screenshot a bunch of them and
6:26it'll just kind of optimize from a
6:28distance perspective how I get to places
6:30or like should I order this piece of
6:32food this piece of food? Like little
6:33simple
6:34>> Yeah, like yesterday I was
6:35on the subway and I took a screenshot on
6:37Google Maps and sent it to like Claude
6:39and in just the chat and I was like,
6:41"Yo, can I pay Apple Pay on the subway?"
6:43Cuz I'm from Australia.
6:44>> Yeah.
6:44>> And he was just like, "Yep, you can."
6:45And I was like, "Wait." Like just little
6:46questions and answers.
6:47>> Little simple things you still should
6:49use chat.
6:50>> Yeah, yeah, agreed. Cuz like sometimes
6:51using an agent to ask like how many
6:53calories in a banana is like killing a
6:55fly with a bazooka. It's like overkill.
6:57Like you don't need an agent for that.
6:58So chat is still like very useful for
6:59me. Like I just got caught on my phone
7:01and like ask the questions like that
7:02throughout the day. But any meaningful
7:03work
7:04straight to the agents.
7:05>> Yeah.
7:06>> So I mean like a good way to illustrate
7:08kind of this old way of working with
7:10chat versus this new way of working with
7:12agents is like a marketing example cuz I
7:14obviously like I used to run an ad
7:15agency. It's kind of used to be my bread
7:17and butter. And
7:19if I was going to create a marketing
7:21campaign, right? Let's just say you're
7:22like doing a Christmas in July campaign
7:24for your e-com brand.
7:25The old way was you probably have the
7:27brief in Notion or whatever project
7:29management tool you're using. You would
7:31copy that Christmas in July brief, paste
7:33it into ChatGPT. "Yo, we're doing this
7:36campaign. Like create me some ads. Write
7:37me the email copy."
7:39You would then like take that copy and
7:40like paste it into Klaviyo or then like
7:42give it to your email designer, to get
7:44the ads, and then like go into some
7:45other tool and make them.
7:47And you'd copy the content from chat
7:49into your content calendar, paste it,
7:51and then you would have to go into like
7:52ads manager and like manually input it
7:55all as well, like once the ads are done.
7:56>> costs and a ton of like small
7:57optimizations. It's It's going to two,
7:59three, four, five different people.
8:01>> And it's back to the thing about like
8:02more fulfilling work as well. It's like
8:04when I used to have to upload If you
8:05were uploading like a hundred ads to
8:06that manager, like the UI in Meta Ads
8:09Manager is so awful. It would take you
8:11like an hour of like just data entry,
8:13like putting them into Ads Manager and
8:15switching off all like the creative
8:16enhancements and all the all the
8:17[ __ ] So, the new way of working
8:20would literally be you just tell Claude
8:21or whatever agent you're using,
8:23you click kick off the Christmas in July
8:25campaign, and it would go into Notion,
8:27it would find like the the campaign
8:29outline about like the plan and like
8:30what specific discount you're going to
8:32give and the offer and stuff. And it
8:33could go into Shopify, like pull in
8:35actual product images. It have like your
8:37brand guidelines so that email would be
8:39like nice and branded.
8:40And then
8:42it would like
8:43write the ads. It can even use a tool
8:44like Higgs Field or an image gen to
8:45create the static ads. And then it might
8:47just give you back like five emails at a
8:49hundred ads. And you pick like your
8:50favorite email and like top ten ads.
8:53And then you approve it and it will just
8:55literally go into Clavio, set it up, and
8:57then go into Ads Manager, set it up, and
8:59like it's done.
9:00>> Yeah. And it And it could take it there
9:02and just leave it as a draft. And
9:03obviously, you can by hand optimize it
9:06before you click go. And for everybody
9:07out there listening and watching, just
9:09to give you context, like we're talking
9:10very macro right now. We are in this
9:12episode going to actually show you live
9:14how to do stuff like this. Because
9:16there's a lot of There's a lot of stuff
9:17like in between there to make sure that
9:18the piping is set up properly to get
9:20that end result. But yeah, man, I just
9:22feel like all of like the little small
9:23work is going to be absolutely done if
9:26and when you set up the proper piping.
9:29It really just comes down to the proper
9:30piping and systems that you have
9:32underneath all of this.
9:33>> Yeah. Yeah, pretty much. So, you've
9:35given the agent a goal. It's It's done
9:37this amazing work for you and given you
9:38the output, but like what has it done?
9:40Like what happens in that mysterious
9:42agent step where it's going and doing
9:43the work? I think that it's really
9:45really important to understand like
9:46what's going on under the hood here. I'm
9:48just going to do it at a high level cuz
9:49I'm I'm not a technical person myself
9:51and I'm not going to lose anyone with
9:52like
9:53going into like the actual software
9:55engineering part of it. But, I just want
9:57to touch on this loop that agents go
9:59through. So,
10:01when you hand an agent a task,
10:04it's literally just going through three
10:05steps. Observe, think, act over and over
10:08and over again in a loop until it has
10:10completed the task and then it hands you
10:13the final result.
10:14So, when I say observe, think, act,
10:16you can think of it as look at the
10:18situation,
10:20decide the next step,
10:21do it. Look at the situation, decide the
10:23next step, do it over and over and over
10:25again. Optimizing, optimizing,
10:26optimizing.
10:27And then finally, it will output the
10:30result when it determines that the task
10:31is done. And then how it determines when
10:34it should stop going in loops is based
10:36on what you've told it in the goal.
10:39So, if you for example said like
10:41to an agent as a simple example,
10:44research the top 10 business podcasts I
10:45should go on and put it in a 10-page
10:49PowerPoint deck.
10:51The agent will not stop going in loops
10:53until it has 10 business podcasts in a
10:56PowerPoint to give to you. So, if you're
10:57really vague with like
10:59what you want the final output to be,
11:01then um
11:02>> It's going to be a little bit more
11:02subjective based on the agent. It's not
11:04going to be exactly what you want.
11:05>> Yeah, exactly. So, it's important that's
11:06one thing we touch on is
11:08clearly defining like what done looks
11:09like.
11:10>> Yeah.
11:10>> But, you know, when you're giving it a
11:11goal, it's just going in these these
11:12these three steps. Look at situation,
11:14decide the next step, do it.
11:16And I thought it would be cool just to
11:17like sort of illustrate this loop is to
11:19just give a simple example into Claude
11:21and I can show you this loop going in
11:23real time.
11:24>> Before we we we show that to everybody
11:25out there, I think a key thing to touch
11:27on just cuz guys, I want to make sure
11:28that you're following along as we go. In
11:30that observe loop that we're popping up
11:32right now, you see the context and
11:34skills. We are going to get very, very
11:36deep into how to set up the context and
11:39the value of the skills.
11:41>> Yeah. Let me give this little example to
11:42Claude so I can show you this loop in
11:43action.
11:45Okay, so I'm in Claude code right now
11:47and we're going to just paste this in.
11:48>> Welcome back, Rem dog.
11:50Before you dive into this right now,
11:52guys, this is going to be awesome. You
11:53guys are going to see the whole entire
11:54loop. Just so people have like context,
11:56I don't even know if you even maybe want
11:57to zoom out from a macro perspective or
11:59we can do it after the fact of just like
12:01driving around Claude and just telling
12:04them like this is code, this is co-work,
12:06this is chat. Let's just make sure that
12:07we do that before the end of the
12:08episode.
12:09>> Yeah, yeah, of course. And like I I want
12:10this episode to be core principles
12:13timeless. So I'm not going to go too
12:15deep into like right now the best model
12:17to use in Claude's Opus 4.8 or Fable 5.
12:19But in 2 months' time, that's going to
12:21be different. But these core principles
12:23that I'm going to cover today about like
12:24how to make an agent actually useful and
12:26work for you and like how to build
12:28agents essentially.
12:30That's not going to change for months.
12:31So for this example, what I've told it
12:34is I've said, "Please give me a report
12:35of my best performing Instagram videos
12:37and a full breakdown of the angle, the
12:39hook, topic, title hooks and why the
12:42video worked." I've given it my
12:43Instagram handle and I've said create a
12:45HTML report. So that's just going to
12:47keep going until it's done an analysis
12:49of my best performing videos. I could
12:51have even given it a number like 10 best
12:52performing.
12:54And then it's going to deliver the HTML
12:55report and that's when it will stop the
12:57loop and give it back to me. So
13:00the first act step that it decided to do
13:03was go and search my computer to see if
13:05it can find more information about like
13:07title hooks and hooks and some of the
13:08things I've talked about.
13:10And it's then found that title hooks are
13:12exactly what I need. So now it's gone
13:14back to the observe step. Now it's
13:16evaluating the situation after the first
13:18loop. So
13:20second loop round, now it's going to go
13:23and read it's it's found some files.
13:25Now it's evaluating the situation.
13:26Great, I've found more files.
13:29What should I do next? Let's look at
13:30these files. So you can see it's read
13:31these two files.
13:33Back to the observe step. Great, this is
13:34a goldmine. It's a complete data set of
13:3660 reels matched with play counts.
13:40And then it has I mean I think maybe I
13:43have my video my real performance
13:44already downloaded on my computer from
13:45another project that it's found. If it
13:47didn't find all these 60 reels with
13:50data, it would have instead of that it
13:53would have gone on the loop and then
13:54probably called it told it to go search
13:55Instagram essentially.
13:57So
13:58second loop round it's it's read the
14:00files. Now it's like excellent, this is
14:01a goldmine.
14:03It's read some more files.
14:05Another loop.
14:06Now it's gone I have a complete real
14:08data set all 60 reels matched with play
14:10counts, engagement rates, dates, patent
14:12tags, all that stuff.
14:14And then it said the one field that
14:16genuinely isn't in this data is the
14:17spoken hook first three to five seconds.
14:19That's the bottleneck.
14:20>> Yeah.
14:20>> So now it's gone it's thinking it's like
14:23well we need to go find that. So now
14:25it's going to run a command and it is
14:27going to go launch a sub agent which
14:29we'll touch on later.
14:31Which isn't as
14:32crazy sci-fi as it sounds. It's just
14:34basically it's launching another version
14:36of itself to just go check.
14:38And then
14:40it's back in the observe step. Great,
14:41now we've got it. The data is rich and
14:43complete. Spoken hook retrieval is
14:45running. Let me build the full report.
14:47And now it's running another act step to
14:49build the report. So
14:51you know, we've got this loop here that
14:53the agent runs through. And a
14:56application or or a program that
14:58facilitates that agent loop just allows
15:00that agent loop to run is known as an
15:02agent harness.
15:03And then all of like the top platforms
15:06you you see online now that you hear
15:07about like Claude code, Claude code
15:09work, code X, Manas, Perplexity
15:11computer,
15:12those are all just different agent
15:14harnesses. They're just different
15:15applications to facilitate this loop.
15:18So when people talk online about I built
15:21an agent that does this.
15:23That would imply they've built a
15:26competitor to these like they built
15:27their own Claude code. They're not
15:28building
15:29>> did not.
15:30>> Yeah, yeah, they're not they're not
15:31building a loop like the better words to
15:32describe it is they tuned an agent. So
15:36out of the box they all go through this
15:37loop, right? But we can actually we have
15:40four levers to pull to make these agents
15:42better, which is kind of like the whole
15:43concept of this episode. We're going to
15:45just do a deep dive into each of these.
15:47But at the observe step that looks like
15:50context, which is the information you
15:51can give it.
15:52And then skills, which is the processes
15:55that you've given it that it can follow.
15:57On the think step is the LLM model. So
16:00obviously these agents that step where
16:02they're you know deciding what to do
16:05the brain behind that is the large
16:07language model, which you like the large
16:09language model is like
16:12GPT 5.5 or Opus 4.8 like the actual
16:15model itself.
16:17And then on the act step, which is
16:19giving it more tools. So like out of the
16:21box the Claude co-worker Claude code or
16:23most of these agents they have like a
16:25few tools. One of them being like a web
16:26search.
16:27But if you want to then have the agent
16:30like handle your email inbox, you need
16:33to tune that act step and give it more
16:35tools.
16:36>> Yeah, I think what's really interesting
16:37about that is it's almost like a sports
16:39or like F1. Like you could have a
16:42Lamborghini and then you can have a
16:44Volvo. Those let's just look at those
16:46and make the analogy that those are kind
16:47of the LLM models.
16:48>> Yeah, yeah.
16:49>> could beat the Lamborghini if the
16:53context, the skills, and the tools are
16:55really [ __ ] tuned and super super
16:57dialed.
16:57>> If you put like a V12 engine in there in
16:59the Volvo.
17:00>> I'm I'm not a car guy. So I don't know
17:01if that was I don't know if that was the
17:01right analogy, but it's just like
17:03>> Like this this is like one of the
17:04analogies that I love to give is it's
17:06like to today in this episode, I mean
17:08you've already started, but you're
17:09learning how to drive.
17:11>> Yeah.
17:11>> And all of these different tools on the
17:13market like Manas, Perplexity,
17:16Claude, Codex
17:17you You think of them as all just
17:18different cars. Like Codex is a Ferrari,
17:21like Claude
17:22Claude Code is a Range Rover, Manas is
17:24like a G Wagon. And as long as you know
17:26how to drive, which is like just the the
17:29core fundamentals, which you know, in
17:31driving it's like steering, brakes,
17:33accelerator,
17:34you can pretty much hop in any of these
17:36cars and within 5 minutes like figure
17:38out how to work it. You know, like
17:40there'll be slight differences, like the
17:41indicator might be on the other side or
17:43like one might have another cool feature
17:44like cruise control that another
17:46doesn't, but essentially they're all
17:48just the same thing and as long as you
17:50know how to drive, you can pop open any
17:51of these and and figure it out and start
17:53using it efficiently. Like
17:55>> Hot take.
17:56>> Hot take there. Everyone is just so
17:57bullish on a specific AI model, but
18:00we're going to prove otherwise today.
18:01Let's dive into what does that loop look
18:03like right now on the IG side?
18:05>> Let's check on this on this little
18:06thing. Let's check on old boy. All
18:08right, so we can see here we've got a
18:09nice report here, which is what actually
18:10works on AI with Remi. And you can see
18:12it's analyzed all 60 of my reels with
18:14the top one being like over a mill views
18:17and we can see that it's like
18:18categorized them. It's like kind of put
18:21them into into categories. It's done a
18:23breakdown of each one with like the
18:25title hook like on screen.
18:26>> Yeah, but I think I mean
18:27>> And if you guys look at this too, I
18:28mean, you can take direct inspiration
18:30from it. That's like you see the right
18:31versus wrong way to use Claude. You can
18:33do another video that says right versus
18:34wrong way to do X. It's saying don't ask
18:37Claude for business advice. Obviously
18:38contrarian in nature, you could just say
18:40don't ask X or don't do X to kind of
18:43>> You can literally just take all of these
18:45to make all of your new ideas.
18:46>> Exactly. I mean, this was just a report
18:48to illustrate this example. Straight
18:49away I can think of a use case here like
18:50if you do content and brand deals, like
18:52you might want to report like that
18:53generated four videos you did for the
18:56brands, they can have like a report of
18:58how the video performed for them as a
18:59company. Like if you're doing brand
19:00deals and stuff.
19:01There's so much stuff you can do here,
19:02but you can see it's just gone through
19:03analyzed all my videos and like pick
19:05them apart like why it worked, what
19:06didn't. We can see here all 60 reels
19:08ranked by plays with the date, title,
19:11hooks,
19:12the durations, and then throughout this
19:14process, right? Like I started to
19:16illustrate before how it's going through
19:18loops about like, you know, it's
19:19observing, it's decided to go find some
19:21files, it's fed that back in, now it's
19:23decided like, all right, we'll go do
19:24this. It it it couldn't find some data
19:28like the first spoken 5 seconds of the
19:31videos. So then it had to like call like
19:33make it another couple loops to go find
19:35that.
19:36And hard to say, but I reckon this
19:37probably would have ran through about
19:38like 20 or 30 loops before it decided
19:40that the task was done. And you can see
19:42here it worked for like 20 minutes or
19:44whatever it was. So it just it it will
19:45just keep going until it's decided like,
19:48all right, this is done, I'm ready to
19:49give it to Remi.
19:50>> Yeah, that's amazing. Everybody out
19:52there listening, you know, if you're
19:53running a company, if you want a report
19:55at 7:30 a.m. in your mailbox every
19:57single day, a beautiful PDF looking at
20:00social media analytics, e-commerce
20:02analytics, anything, yeah, we're going
20:04to walk through how to do this, but it's
20:05crazy how granular this can get on, you
20:07know, that individual prompt. Quick
20:10>> Second, Remi has shared so much in this
20:11episode. All of his principles and
20:13frameworks we put into a free PDF
20:15playbook you can download up below, link
20:17in the description. Enjoy the rest of
20:19the episode.
20:20>> It's really about what's what's
20:21underneath this and obviously the
20:23piping, so let's get into how to
20:25actually, you know, you call it tune to
20:27to optimize the agent to make it useful,
20:30so why don't you kind of walk us through
20:32those three big pillars and we're going
20:34to deep dive into each.
20:36>> 100% so
20:38a raw like agent out of the box, so if
20:40you open up Claude code or or co-worker
20:41or any of these popular agents on the
20:43market and just like give it a task,
20:44like it's
20:46they're really capable, but it's
20:47probably not going to do it in the way
20:48you like and you're probably going to
20:50get an output you're not happy with. And
20:51that's why a lot of people like start
20:53using these tools cuz it's all the rage
20:54and then they just quit cuz they're
20:55like, oh, it didn't work for me.
20:57In order to actually make the most out
20:59of these agents and make them actually
21:00useful, you need to treat setting them
21:02up like onboarding a real employee.
21:05So you can think of agents as like
21:06really really capable strangers. Picture
21:08that dude, you know, like we've all got
21:09that friend that's like really, really
21:11smart, like good at solving problems,
21:12and will just like keep working until
21:14like the task is done.
21:16You now have the opportunity to hire
21:17this guy into your business for $100 a
21:20month in salary.
21:21>> $1,200 a year, sign me up.
21:23>> And and he has infinite bandwidth. He
21:26can just
21:27He like like there's no limit to what he
21:29can do, but no matter how great he is
21:31and smart and capable,
21:33unless he fully understands you and your
21:35business, or if you're an employee, your
21:36role and what you do,
21:38unless he has the tools he needs to
21:39actually do the work, like an email
21:41account, a browser, a computer,
21:44and lastly, unless he's trained on like
21:46your secret sauce that makes your
21:47business unique and your processes,
21:50he's not going to be very useful at all.
21:52>> All right, so let's actually onboard
21:54this hire, and you mentioned context
21:56tools and skills. Give us full context
21:58of how you look at onboarding the hire.
22:00>> Like I said before, this this really
22:02capable stranger you've now got working
22:04for you, he's not going to be able to do
22:05very much unless he fully understands
22:07you and your business, has the tools to
22:09do the work, and then knows how you like
22:11the work done.
22:13And you do this through context, tools,
22:15and skills. So, context being like
22:17information to the agent, like what you
22:19do, your business, like your brand
22:22voice, things like that.
22:23Tools being like, you know, a computer,
22:25a browser, Gmail, Notion, Slack. And
22:28then skills being like your SOPs, like
22:30your special sauce, like literally like
22:32documents like you give a VA, and then
22:35if you get those three things right,
22:36that's how you like properly onboard the
22:38stranger, and then you have the best
22:40employee you've ever had.
22:41>> Sounds amazing. Let's go super, super
22:44deep into each one of them. So, we have
22:46context, tools, and skills. Let's double
22:48click on context and tell us exactly
22:50what you mean by context.
22:51>> Yeah. So, context is literally like
22:54everything that the agent should already
22:56know before you give it a task.
22:59And for me, this looks like things like
23:01about me, so all the information about
23:03like me as the user, cuz I want him to
23:05like I want my my new employee to
23:07understand like everything about like
23:08who I am and what I do. Even some things
23:10about like my story and like my come up
23:12and things, cuz then it can use it if
23:14it's ever writing like a landing page
23:16copy or like a sales letter script, it
23:18can reference like real things in my
23:20life.
23:21It also looks like a
23:23document about your business
23:24information. A like a maybe another
23:26document of like your offer catalog,
23:28where you talk about like all your
23:29services and offers and their prices and
23:30how they link together, like upsells,
23:32downsells. Another one might be like an
23:34ideal customer profile, which is one of
23:36my like most important ones, so it like
23:38knows who your business like targets.
23:41And quite literally,
23:44if you were onboarding this new
23:45employee, right? Like a real employee,
23:47you'd probably like give them PDFs or
23:49like some sort of documents, usually
23:50like Google Docs, Word Docs, that teach
23:52them all of this stuff or like the key
23:54information you want them to know. You
23:56You are literally doing the same to the
23:57agent. So like just like, you know, you
23:59get a VA get a new VA, you would like
24:01give them all these docs, you're
24:02literally giving these docs to the
24:03agent. And these are in the form of
24:07markdown. So,
24:09for the listeners that don't really
24:11understand what markdown is, markdown's
24:13like
24:14it's it's basically
24:15one step above a TXT file. So, .txt text
24:19file is just literally plain text. No
24:21headings, no bold, like nothing, it's
24:23just text.
24:24Markdown is literally just one step
24:26above that. It's just a plain text file,
24:28but you can do like headings,
24:30boldings, like bullet points.
24:32>> Simple formatting, stuff like that.
24:34>> Yeah, it's just like a slightly more
24:36formatted text file. And you know, if if
24:38any of you guys out there use like a
24:39tool like Notion, the Notion Docs or
24:41like ClickUp, the ClickUp Docs, they're
24:43all in markdown format. Just super
24:45simple, like headings one, twos, and
24:47threes, text, bullets, and bolding, and
24:49like italics, just the super basic
24:51formatting.
24:51>> And to give you guys further context,
24:52it's basically like a simpler .doc file.
24:54So, it's a .txt, .doc, it's somewhere in
24:56that universe.
24:57>> Exactly. And the reason like why
24:59markdown is so important is you could
25:02theoretically, let's just say you're
25:03building you're you're following along
25:04and you're you're trying to build out
25:05your own agent, you could just give them
25:07all these in a .docs file, right? Or
25:09PDF.
25:11However,
25:12it it's a lot harder for the agent to
25:14access the information. Like if you give
25:16an agent your ideal customer profile as
25:17a markdown file versus a PDF, with the
25:20PDF it's going to have to like
25:22unpack it and like try to extract the
25:24text cuz it's just the way that the PDF
25:26file is like built.
25:27It's hard for the agent to like ingest
25:29the information, whereas
25:31with markdown, it's just like that
25:33straight ingested. It can just read it
25:34really really easily. And we want to
25:35make things easy for the agent cuz that
25:37just means it's things are quicker for
25:39you and things are cheaper in terms of
25:40like cost for the model running. So,
25:42markdown is just the best way to be
25:44giving information to your agent.
25:46>> people context as well, how many .md
25:47files do you have on your computer?
25:49>> Well, like hundreds.
25:50>> Hundreds?
25:51>> Hundreds.
25:52But like so they these
25:54you're literally building out like
25:56markdown files for key areas of like
25:58information that you want to give this
26:00agent.
26:00>> Yeah.
26:01And as far as that goes, too, you can
26:02get would you just say the more granular
26:05the better?
26:06>> Not necessarily.
26:07I don't really think there's like a
26:08right or or wrong way to do it. It's
26:11more like I I just build the markdown
26:13files for me almost like how I would
26:14want them formatted. Like I've just like
26:16literally got like a about me, like a
26:18business info, brand voice, ideal
26:20customer profile, like offer catalog. I
26:23could put them all in one doc, but then
26:25it's just like I don't know. I I almost
26:27just prefer them having having them
26:28separate.
26:29>> Yeah. And we're going to get into folder
26:30structure as well, too, so everybody
26:31understands kind of from a macro
26:33perspective how this brain looks and how
26:34it kind of all flows. I think for the
26:36people out there listening, you know,
26:38we'll get in, you know, what is the
26:39first thing that you can do to kind of
26:40start your journey.
26:42>> Yeah.
26:42>> I feel like like the about.md file is
26:44like a pretty clear one. I think it's
26:45like
26:46>> Greatest place to start. And
26:48the thing you need to understand is
26:49these markdown files you're building are
26:50like assets. They're the assets for you
26:53and your company. So like really like
26:55spending the time to build these docs
26:56like really really well. Like if you
26:59just like take the time to build a
27:00really well thought out about your
27:03business doc. And that like the ideal
27:04customer profile, they're they're assets
27:06for your company because if you rush the
27:08ideal customer profile doc and it's just
27:10like a bunch of AI slop that you got
27:12ChatGPT to put together and you haven't
27:13read over every line, there might be a
27:15line in there that says like
27:17my ideal customer like if you're a skin
27:19care brand, a line in there in your
27:21ideal customer profile might say their
27:22problem point is like
27:24red skin or something. And if you don't
27:26check over that, now in all your
27:28marketing copy, you're going to have
27:30like it targeting towards the angle of
27:32red skin and if that's not something
27:34that your product solves,
27:36then
27:37that's like a that's an issue, you know.
27:39>> around the MD files is probably
27:41you want to minimize the subjectivity in
27:43the MD files, make them all objective
27:44like stuff that's kind of black and
27:46white. Like I definitely would use this
27:48as an input to make decisions.
27:49>> Yeah, yeah, agreed. And I and I get
27:51Claude to build all these files for me,
27:53but I read over every line before I like
27:55give it the okay.
27:56>> Guys, I got a little cheat code for you,
27:57too.
27:58On the about that MD file, like the
28:00craziest thing about all this is like
28:02I'm not saying have Claude do it, but
28:04you can say to Claude, I want to build
28:06an about that MD file. How should I you
28:08know, how should I process this? How
28:09should I think through this? And that's
28:11where you I do think you can use
28:12something like chat cuz you can kind of
28:14go back and forth and build it.
28:15>> Yeah, that's like a a perfectly fine use
28:17case for like the chat models. Like
28:19that's how I built out all of these. I
28:20literally said to Claude like
28:21in a new chat, I was like, "Yo, I want
28:23to you to build out a bunch of markdown
28:25files for me. Like a an about me and
28:27ideal customer profile." Just listed out
28:29all the ones I wanted and I just said,
28:31"Interview me like
28:33to extract the information from my brain
28:35to build these files." And I sat there
28:37like using Whisper Flow answering all
28:39the questions for like 3 hours.
28:41>> Just riffing and then Whisper Flow just
28:43cleans it all up.
28:43>> Yeah, it just like sucks it all out of
28:45your brain into these files that are now
28:46assets for your company that you can
28:48then give to your agents. And like one
28:49of the key distinguish like the key
28:52distinguishments to make is when you're
28:54using these chat models,
28:55they're sort of like a black box. Like
28:57everyone
28:58always goes on about like my chat GPT
29:00knows so much about me. I don't want to
29:01move into Claude because
29:03like Claude doesn't know me. Chat GPT
29:05knows me so well.
29:07And it's like you're right, but what
29:08it's doing is
29:10it's
29:11like got pretty much just a markdown
29:13file like memory.md or like about
29:16user.md in the background. And
29:20Chat GPT's building this file on your
29:21behalf without you knowing and you can't
29:23access this file. So if you're like
29:24starting a Chat GPT chat and you say in
29:26there like
29:28I'm I've just moved to New York City.
29:29Like what are some things like good
29:31cafes near me?
29:33Chat GPT in the background will take out
29:35that snippet of like I just moved to New
29:37York and save it in its markdown file.
29:39And then now like it knows you've moved
29:41to New York and you just think that this
29:42model's like almost trained on you now,
29:45but in reality it's just a markdown file
29:47where it's choosing what to save. And
29:49the other thing about that is
29:51you can't control what it saves as well.
29:53Like your Chat GPT like knows everything
29:55about you. And then like you know,
29:56you're writing like a landing page.
29:57You're asking for advice on Chat GPT
29:59like
30:00Hey, how how do I break up with my
30:01girlfriend? And then you're writing a
30:02landing page and then it's like all of a
30:03sudden the landing page copy is about
30:05you breaking up with your girlfriend.
30:06It's like
30:07you can't you you have no control with
30:10these chat models about like what it
30:12knows and what it saves. Like it it does
30:14get mixed up
30:15>> sometimes.
30:15>> Exactly. So you want to make sure that
30:16you have full control. You want to own
30:18that context. And that's like I'm
30:20talking about just in the format of
30:22these markdown files. And then you can
30:23they're assets that you own now and you
30:25just bring them to whatever agent you're
30:26using. So like I've got my folder set up
30:29which we're going to get into with all
30:30the context in there. And then I can
30:32just open up that folder in Claude code
30:35or Codex or Manas and like straight away
30:37it knows everything about me. It's like
30:39it doesn't matter like I can just switch
30:42and stay sort of like mobile.
30:44>> Guys, real quick, everything that Remy
30:46is teaching you in this video is
30:48absolutely free. No gatekeeping, as
30:50always. But for the people who want it
30:52implemented, Remy made a plugin that
30:54builds this entire system for you,
30:56around your business and your life. The
30:58loop, the folder structure, all of it. I
31:00am not technical. I did it myself. 90
31:03minutes, one sitting, done. Sign up for
31:05the newsletter below and get 50% off at
31:08launch. Enjoy the rest of the episode.
31:10>> So, the first step for onboarding agent
31:11is like building out these context
31:12files, right? And I'm going to show you
31:14like what this actually looks like. So,
31:16I'm going to open up Claude Code again
31:17here.
31:19We're going to do a new session.
31:21Nice and clean.
31:23Hey, there we go. And I'm got my OS
31:25folder open now, which we'll touch on
31:26later. But, the way that these agents
31:28work now is a lot of them run like
31:30locally on your computer, as opposed to
31:32like the cloud.
31:34And
31:35what locally on your computer means is
31:36it's like instead of having like a
31:37project on your ChatGPT account, your
31:40that project folder now lives on your
31:41computer. Yeah.
31:43And there's a lot of benefits to that,
31:45which I'll I'll touch on later. But, I'm
31:47just going to open up this one I've
31:48created now, called demo. So, this is
31:51just a brand new folder. You can see
31:52it's completely empty.
31:53>> And that's just you just created an
31:54empty folder on your desktop.
31:56>> Yeah. So, it's desktop, empty folder
31:58called demo. I'm just going to open that
31:59up.
32:01So, now you can see in Claude Code,
32:02we've scoped the agent to just work in
32:05that folder.
32:07Just like you know when you have a a
32:08project in ChatGPT, it's like any of the
32:10chats you do in that project, it's like
32:13scoping ChatGPT to just work there.
32:14>> So, now it's a new blank canvas, and
32:16you're and you're deciding what inputs
32:18go into this folder.
32:19>> Exactly. Exactly. So, you can see, just
32:22to demonstrate this the importance of
32:23context, if I'm opening up this brand
32:25new folder,
32:26so, I've been working with Claude Code,
32:27right, for for
32:29like 8 months.
32:31And
32:32if I ask like, you know, who am I?
32:37So, you can see like when I've asked it,
32:38there's no markdown files for it to look
32:40at. So, I've asked it how am I? I've
32:42been using Claude for 8 months, it
32:44should know, but it doesn't. It's
32:45because all the context you own in those
32:47files. It's not a black box.
32:48>> So, someone has an alter alias, they can
32:50they can just create different alter
32:51aliases in in in different folders.
32:53>> Yeah. Yeah, essentially. So,
32:55you can see here it tried to read the
32:56memory.md mark file markdown file, but
32:59it doesn't exist. So, you can see here
33:00it's just like the file file doesn't
33:01exist.
33:02>> went through that loop. It was
33:03searching, but there was no kind of raw
33:04data to extract.
33:05>> nothing for it to to do.
33:07>> So, how did they know the email and the
33:08likely name?
33:09>> Cuz that was just from my Claude
33:10account. It must just have access to
33:11that. That's literally just cuz that's
33:13the email associated with my Claude
33:14account and name. Actually, no, no, no,
33:16actually it's it's pulled that from my
33:18computer because this agent lives on you
33:20know, it's got you've got a username on
33:21the computer. That's where it's gotten
33:22it from.
33:23>> Okay, so what does this look like when
33:24you're on steroids? Let's dive deeper
33:26into kind of doing these MD files.
33:28>> Yeah. So, let's let's continue on
33:29boarding our agent, right?
33:31So, the most important markdown file,
33:33like you know, you've got a markdown
33:35context files. The most important one
33:37when you're dealing with agents is
33:38what's called an agents.md file.
33:41And I mean, Claude have just decided to
33:43be difficult and in in their systems
33:45they call it a Claude.md, but in pretty
33:47much all the other harnesses you use
33:48like Codex and Open Claude and stuff,
33:50it's called an agents.md.
33:52And essentially what happens is is every
33:54time you're like starting a new session,
33:57and by new session I just mean like if
33:58we go back to that goal to result thing.
34:00It's like every time you're going to
34:02like give it a new goal, you're kind of
34:03going to click like new session. Same
34:04with you starting a new chat in ChatGPT.
34:07Each session has an amount of like
34:08space. It's kind of like you can think
34:10of like the agent's like brainpower. And
34:12as the session starts to like fill up,
34:15the
34:16you have what's called context rot,
34:18which is basically like have you ever
34:19been in a in a chat on ChatGPT and you
34:22it's like so long you've been in there
34:23for hours and it starts to just like go
34:25crazy. Like it literally just starts to
34:26just like hallucinate, forget things.
34:28So, what what's happening there is every
34:30time you start a new session like a
34:31fresh like chat window, I guess, it's
34:33got like a certain amount of of I guess
34:36capacity, you could call it.
34:37Like free for the actual work. And what
34:39your session is probably look like is
34:40something like this, where they're just
34:42like
34:43like so like so much information that
34:45the stuff you started telling it at the
34:46start starts to fall out the top.
34:48>> V1 Mark bro. I'm not on V1 Mark anyway.
34:50I mean, I think 99.999% of the world is
34:53in probably V1 Mark 1 endless chat, to
34:56be frank with you.
34:56>> Yeah. Yeah, probably. So, every time you
34:59start a fresh session, the way that this
35:00agent like knows about you, same way
35:02that ChatGPT's memory works, is if you
35:04start a new chat, in the background,
35:05it's started the chat with that
35:08memory.md loaded into it.
35:10But you can't see that, right? So, you
35:12give it a prompt, like uh write me an
35:14email in ChatGPT, and to you it just
35:17shows as like that one prompt you've
35:18given ChatGPT, but in the background
35:20it's also loaded in that memory markdown
35:22file. So, instead you've got like
35:24like a hundred lines of information
35:26about Mark, and then write me an email.
35:29So, it's got all that context in it.
35:31Same thing happens with agents, right?
35:32So, we've started a fresh session here.
35:34What this actually looks like is any
35:36file, any markdown file in that folder,
35:39titled Claude.md, gets automatically
35:41loaded in. So,
35:43your Claude.md, or agents.md,
35:47that's like your your North Star
35:49markdown file. That's the one you need
35:51in every single project you're working
35:52in. So,
35:54to illustrate my point, I've created
35:56this folder here with a Claude.md in it,
35:58right? That I've pre-prepared. And we
36:00can actually just open this up, so you
36:02can actually see what a markdown file
36:03looks like. And I got Claude to build
36:05this for me.
36:06So, we can see
36:07Remy's AI operating system, who am I?
36:09Remy's a a founder of AI with Remy.
36:12The business, the brand voice, the ideal
36:14customer, the offer, the values, how I
36:17like to work, preferences.
36:19This is like our, you know, North Star
36:21context file. And that's the one that's
36:22going to load in to each new session by
36:24default.
36:25>> To give macro context on how you're
36:27thinking.
36:27>> Yeah, to give macro context essentially.
36:29So, that same example, right? If we go
36:31back into Claude here, we just asked it
36:34in that folder when it was empty, "Who
36:35am I?" It had no idea. Now, if we go new
36:38session,
36:39same folder, but it's now got that
36:41Claude.md in it, and we can say, "Who am
36:43I?"
36:44>> Hey-o.
36:46>> And you see here,
36:48now it knows
36:49exactly who I am. It knows exactly like
36:51what I do, my background, the brand, the
36:53three products.
36:54So, that's like a great illustration
36:57there. Do you Do you see how how it goes
36:59from knowing nothing about you to now
37:00having that context file? And it knows
37:01everything.
37:02>> That's why it's so important to be lean
37:05and intentional on what files you upload
37:06obviously in there.
37:07>> Yeah. Yeah, exactly. So, that's that's
37:09your pinnacle markdown context file.
37:11That's like the first one that you need
37:13to build.
37:14The The part about the Claude.md made
37:16sense, right? That gets slotted in, and
37:17it's like the agent wakes up to do a
37:20task, and he's got that loaded into his
37:22brain, right?
37:22>> Yeah, and that's something that's you
37:24should constantly be updating if
37:26something big in your life changes,
37:27correct?
37:27>> Yeah. Yeah, correct. Yep.
37:29So, you're probably like wondering now,
37:30"Yeah, yeah, I get that. Claude.md,
37:32Agent.md, where do the other ones fit
37:34in? Like the about me and the ideal
37:36customer profile? Like where do I put
37:37those?" So, those ones, the other
37:40markdown files, they don't get loaded
37:42into that session window by default. So,
37:44we need to We've got our Claude.md that
37:47gets loaded in by default, and then we
37:48need to tell it in that that it will
37:50always read, "Hey, go check here as
37:52well." So, then what I like to do, how I
37:55have this set up, is we've got our
37:58Claude.md here,
37:59and I'm going to paste in these other
38:01ones I've pre-prepared.
38:03So, now you can see we've got a couple
38:05of other files taking shape here. So,
38:08we've got our Claude.md,
38:09we've got this context folder with all
38:12these markdown files in it, my values,
38:14my offer, my ideal customer profile,
38:15brand voice, about me,
38:17and then this memory.md, which we'll
38:19touch on in a second. So, by default,
38:21Claude is any any project you open up
38:23Claude in
38:25to work, it's always going to read that
38:26Claude.md. And then we're going to point
38:28to the context folder in the Claude.md.
38:31So, you can start to like
38:32compartmentalize and like modulate your
38:34your project, if you will.
38:35>> Making like a whole like a like decision
38:37rubric, like a decision tree of, you
38:39know, if it goes here, it opens up this
38:40folder and stuff like that.
38:42>> Exactly. And then this is like all these
38:43agents you build, they're super like
38:45modular. It's like
38:46you hear other ones really popular like
38:48Obsidian is a really popular thing that
38:50people give to their agents as like a
38:51second brain.
38:53Essentially like
38:54there's loads of different like
38:55frameworks like that. And instead what
38:57they're doing it like if if you had this
38:59folder and you wanted to use Obsidian
39:01instead, you would instead delete this
39:03context folder and you would in your
39:05Claude.md point to the Obsidian vault
39:07instead. That's kind of how people are
39:08doing it. Like you can kind of modulate
39:10the the context.
39:12>> So, in short, you can kind of redirect
39:13it whichever route you want to go.
39:14>> Yeah. Exactly. And it all stems from
39:16this file here. So, now this new setup
39:18we've got is kind of like evolved. So,
39:21we just this the first one was just the
39:22Claude.md. That's probably the most
39:23basic one that if you're trying to like
39:24start using agents for yourself, that's
39:26probably the first step I'd say is just
39:27create a folder and then just build a
39:29Claude.md. Just ask Claude, "Hey, build
39:31this Claude.md file for me and just
39:34include in it like an overview on who I
39:35am and what I do." Just to give it like
39:37that base context. Then V2 is now we're
39:39adding in like this extra folder for
39:41like we're expanding out our our context
39:44here.
39:45And if we open Claude.md now,
39:47now I'm going to point to that context
39:48folder. So, you can see if we open it
39:50up, it's like, "Who am I? My context,
39:52read before any task. Everything you
39:54should know about me and the business
39:56lives in the context folder." And then
39:58it's just giving an overview on the
39:58files.
39:59>> Wow.
40:00>> Load the relevant files in before
40:01starting any tasks. Assumptions are the
40:03enemy. If the answer isn't there, ask
40:05me. And then everything on continues
40:07like normal.
40:08And then the memory.md I'll touch on in
40:09a sec.
40:10>> And then as you keep developing more
40:12folders, you just keep updating. The
40:14most important thing is the Claude.md
40:16file cuz
40:17>> that is going to create the
40:18infrastructure. Like you could have like
40:20marketing, finance, logistics.
40:22>> Yeah.
40:23>> And then underneath that you can have
40:24multiple folders.
40:25>> Exactly. So, when I asked Claude before
40:28like who am I and knew straight away,
40:29that's cuz that file's automatically
40:30loaded in. But if I deleted that
40:32Claude.md,
40:33and then asked it who am I again,
40:35>> Back to dummy. Back to dummy.
40:36>> Well, it's got the context folder.
40:38>> Oh, yeah.
40:39>> If I just deleted the Claude.md and had
40:40left the context one, it wouldn't know
40:42straight away. It would then go through
40:43a loop to think like, "Oh, well, I don't
40:46I don't know who he is. Uh
40:48let me think. I can Let me Let me check
40:50for what folders I've got. Oh, context."
40:52Then it would read it, then it would
40:53load it in. But it's just like it's slow
40:55and it's going to go through another
40:56loop. So, like that's why you have the
40:58Claude.md there.
40:59>> Yeah.
40:59>> Why don't Why don't you for everybody
41:00listening right now just ask it
41:01something very, very basic under the
41:02notion that now you have context based
41:04in like something about any of those
41:07kind of raw folders. I know you had
41:08under there like brand voice, like
41:10anything.
41:10>> Yeah, well, what I What I might do is
41:12I'll just say like, "Write me a
41:14cold email to
41:16>> How about to Mark to come on to come on
41:18Open Residency?"
41:19>> Actually, no. Hey, let me do this. I'll
41:21just say like,
41:23"I'm going to Whisper flow to dummy."
41:25>> Yeah, amazing.
41:25>> Makes more sense.
41:27"Write me a cold email to Mark from Open
41:30Residency podcast
41:31to try and get me a position on the
41:33show."
41:34>> Wow, Whisper flow for the win, baby.
41:36>> Whisper flow is so good. Like
41:39talk about like
41:40agents stacking up the productivity. You
41:42can even stack it on top of the agents
41:43when you start using Whisper flow. Cuz
41:45it's like you can just
41:47just get things down onto the page at
41:49the speed that you can think.
41:50>> For the people out there listening, I'm
41:51going to give you guys a little cheat
41:52code. One of our episodes a bit before
41:53this is of Chris Voss, which is a
41:55negotiation expert.
41:56>> Yeah.
41:56>> You can make some MD files with him to
41:58speak in his language. So, it probably
42:00would be a bit more persuasive on this
42:02email that you're going to send me.
42:03Let's Let's Let's see how good it is.
42:04I'll see your sales pitch right now.
42:05>> Okay, great. So, it's gone through a
42:06couple loops, right? It's got decided to
42:07go and research who you are. It's gone
42:09to go and like find
42:11>> That is me.
42:12>> open residency
42:13And now it's written, right? It's like
42:16literally said here
42:18Like we've been given any other context.
42:19We just said write the cold email. And
42:20it's got cuz it's got our our our Claude
42:23that I'm doing in this context folder.
42:24And that context folder is so rich with
42:26like my values, my philosophies, my like
42:28thought processes on things. And my
42:30offers. So you can see like the subject
42:32lines, right? Operating a 20-person team
42:34without hiring 20 people. Guest idea for
42:36open residency.
42:38But the email right here. Hey Mark, been
42:39working through open residency. The
42:41operator as last line of defense framing
42:43is the realest take on running a
42:44business. Terrible cold email, by the
42:46way. Terrible. But it's basically you
42:49can see here it knows everything about
42:51me. It's like I teach founders to stop
42:52drowning in in tools.
42:54>> This is where we're going to get into
42:55like the additive layers of like an
42:57example is like that negotiation
43:00skill that, you know, how you speak
43:02skill. Like you can get way deeper to
43:03make this thing absolutely flawless. But
43:05clearly it did an amazing job.
43:07>> Yeah, I mean the the key point
43:08demonstrate was, right? That it doesn't
43:10knows about me. It it can write an email
43:12with all the context. Like it doesn't
43:13like the email's terrible, but like and
43:15you can tell it knows about like what I
43:17do in my philosophy.
43:19>> Amazing. You're booked.
43:20>> Let's go.
43:21>> [laughter]
43:21>> We're on. We're here.
43:22>> Okay, great.
43:24>> So and just to take a a step further
43:26just to kind of paint the picture of
43:27where we're going to go with this.
43:28>> Yeah.
43:29>> This could be this email's drafted. It
43:31could also be automatically sent as
43:33well, too, which we're going to get to
43:34that to connectors and MCPs
43:35>> Yeah.
43:35>> too. Just so people know like this could
43:37be fully A to Z find all the business
43:40podcasts
43:40>> Yeah.
43:41>> draft the email and send the email.
43:42>> And then you schedule it. So every week
43:43it's going to every morning it's going
43:44to go find business podcasts, cold email
43:46them to get you book calls, like all
43:48this stuff. It's crazy.
43:49>> You're going to need to dial that email
43:50in a little better for me to say yes to,
43:51but it's okay. Don't worry.
43:52>> Keep going. So that's that's what I'm
43:54going to do now. I want to display like
43:56So we understand like we need these
43:57context markdown files. The most
43:59important one's the Claude that I'm
44:00doing. But then points to the others.
44:03And now like when we open up a new
44:04session, it it it straight away knows
44:06everything about us, right? But, the
44:08next layer of context is memory, which
44:12is still kind of under the umbrella of
44:13context, right? Memory and lessons.
44:16So, off the rip, it knows all this stuff
44:18about us, but what if we ask it to do
44:20something differently? So, you can see
44:22I've got in here this memory.md file,
44:26right? If we open this up,
44:28this is just This is just got memory
44:31learned rules, lessons Rem has taught
44:32me. I check these before every task.
44:34It's got never write formally, just a
44:36couple of like random preferences. So,
44:38the agent like knows everything about us
44:39out the box, but we want it to be like
44:42be able to improve over time. And like
44:43if we correct it, we want it to save
44:44that somewhere, which also counts as
44:46context, right? So, I like to build this
44:49little like kind of
44:51de facto memory system. I think
44:54like when I started putting this
44:55together, it was because none of these
44:56agent harnesses had memory built in. So,
44:58if you said to it like write this this
45:00email right here. If we said like
45:03"Can you please sign off the email with
45:05cheers, Remi."
45:06And like I always want you to sign off
45:08emails with cheers.
45:09>> It's not auto updating on the put it in
45:11the memory.
45:11>> It's not. So, like before last maybe in
45:14the last month, they've a lot of these
45:16harnesses have now added in memory as a
45:17feature. But, before that, you had to do
45:19it yourself. Cuz otherwise, it would say
45:20like if you didn't have that in place,
45:22it would say like, "Yep, got it." But,
45:23then next session in a week time when
45:25you write an email, it wouldn't do it.
45:27>> Show them a memory.md just in case
45:28they're driving the Kia and not the
45:29Lamborghini and maybe that LLM does not
45:31have it built in.
45:32>> Exactly.
45:32>> bit of context.
45:33>> So, we can just show in here like the
45:36memory.md.
45:37>> on the Kia either. I've seen some nice
45:38Kias, by the way.
45:39>> [laughter]
45:40>> So, we've got the the memory learned
45:41rules. These are just a couple I put in
45:42here before like never write formally,
45:44like always use real examples, never
45:46generic placeholder ones. And then, same
45:48way that we're pointing Claude at the
45:51context folder, we're also going to
45:53point it at the memory folder. And I've
45:54just basically said, you know, your
45:55memory lives in this file. When I
45:57correct you, write lessons into this
45:59memory file as a the rule, and then
46:00check it before every task. This is how
46:02you go from good assistant to know
46:04exactly how I like things done.
46:05And I've One little tip is if you use
46:07the at symbol, if you tag a file like
46:09that, then it like it's almost like a a
46:12simlink. I don't know if you know what
46:14that is on a desktop. It's like
46:15>> I don't.
46:16>> If we For example, if I tell it
46:18>> Is it like a a lighter load on the
46:20brain? Like it it zips right into the
46:22file or something?
46:22>> Yep, yep, zips right in the file. It's
46:24it's like as good as the file being
46:25loaded in, if that makes sense. Cuz if
46:27we just tell it like go use the
46:28memory.md and don't add it, it'll then
46:31have to make a loop to go and find the
46:33file and bring it in. But if you just
46:34have like the at, it's automatically
46:36loaded straight in.
46:37>> Guys, we're going on Lamborghini, full
46:38steroids here.
46:39>> Hell yeah. So,
46:42Claude like has this this built in. But
46:44even still, it's still almost a little
46:46bit of a black box. It will save it to
46:47like Claude's desktop app. So, now if
46:50you're trying to work with this folder
46:51in like Codex, or you want to try and
46:53like use other harnesses, you don't own
46:55the context file. Claude kind of owns
46:57it. So, like that's still why in all my
46:59projects, I still kind of have this
47:01memory system in.
47:02So, we can just test this out here,
47:04right? And we can say,
47:06"Please in future, when creating
47:08websites, I want you to always use dark
47:11mode and never ever use light mode."
47:14>> What is that off?
47:17>> All right, here we go. It's done it.
47:18Bang.
47:19All right, great. So, you can see it's
47:20now gone. And instead of add I'll add
47:22this to your memory rules so I never
47:24forget it. It's read the memory.md.
47:27And then you can see it's added it's
47:29edited and it's added this line, "When
47:30building any websites, always use dark
47:32mode, never use light mode."
47:35>> Love that. So, that's that's that's
47:36that's obviously in that auto updated in
47:38that folder that you showed us.
47:39>> yep. So, if we go back here into this
47:41demo folder, and we go into our
47:42memory.md file and open that bad boy up,
47:45you can see
47:46"When building any website, always use
47:47dark mode, never use light mode."
47:49And then now this is like we own that
47:51that preferences, right? This will grow
47:53over time. And like
47:55I've had the question before. It's like
47:57what happens if that just keeps growing?
47:58You've got like hundreds of different
48:00lines in there and they all start like
48:01contradicting and get messy. Like
48:04I don't think that will happen. Maybe in
48:05like 6 months time, but you can just do
48:07a spring clean and just open a session
48:08with Claude and go like
48:10you just audit my memories and just take
48:12out anything that's contradictory or
48:13conflicting and clean things up. So,
48:16that is pretty much like context 101.
48:19Just to recap, we've got the Claude.md
48:21which Claude auto loads in every
48:22session. We want to give it more
48:24context. We add that extra little file
48:26and then point Claude to that folder and
48:29then we've set up our little memory like
48:30self-improvement system as well.
48:32>> So, from a loop perspective, we're going
48:33to start with the about.me, the
48:35Claude.md is going to be kind of the the
48:38macro infrastructure, which is then
48:41going to lead into more folders, which
48:42will obviously get into. And then the
48:43memory.md is just a self-fulfilling loop
48:46that just keeps remembering all of the
48:48preferences and such.
48:49>> Yeah, yeah, exactly. So, now like every
48:52every session
48:54we start
48:56that observe step now has all the
48:58context and skills loaded in. So, every
49:00time we're giving it a goal, straight
49:02away in the first observe step before it
49:03does anything, we've already got like
49:05now I'm in skills we'll touch on, but
49:07now you can see we've really tuned up
49:09that con- context step. So, when it
49:11starts to work on your tasks, it has
49:13like all the information that it needs.
49:15And I think one last thing just to just
49:17to finally like drive home this point
49:20we can run in Claude code this little
49:21command. Don't worry about all the
49:23others, we'll touch on that later. But,
49:24this is a one that comes built in with
49:26Claude. You can go {slash} context
49:29and it will actually pull up that real
49:31bar chart for you. So, you can see. See
49:33here, context window, we've used 6% of
49:36that bar.
49:37>> Are you talking about 6% of like the
49:38memory within that individual chat?
49:40>> Yeah, within that individual chat. So,
49:41you can see it's already like in our
49:43bar, it's used up 6%.
49:46And if we look at like we we've given it
49:48one task here, right? Like please in
49:50future tiny little message and it's 6%,
49:52which is a fair bit. It's almost like
49:54it's a fair bit of of the total space
49:56used. So, you're starting each session,
49:58even if we just did like a blank thing
50:00with no messages sent, you'll always
50:02like have a little bit of your window
50:04chewed up because it's feeding in all
50:06that context before you start any task.
50:08>> That's why you could also argue, going
50:09back to the beginning, is keep the MD
50:11files like lean and cut, you know, only
50:13use what you need to use.
50:14>> Yeah, exactly. Cuz if you have these
50:15like really bloated context files and
50:17like you have let's say you have 100
50:18markdown files, you start a session and
50:20the context window is already like 20%
50:22full. It's like you you want it Yeah,
50:24100% to your point. You want to keep
50:25them nice and lean and like really like
50:27potent with like high-density
50:29information.
50:30>> Amazing. I use WhisperFlow for hours
50:32every single day. I think way faster
50:35than [music] I type. I speak about 220
50:38words a minute and type only 45.
50:40WhisperFlow closes that gap. It's a
50:43voice-to-text tool that works anywhere
50:45on your computer or phone. You talk and
50:47clean formatted text shows up wherever
50:50your cursor is. All right, let me show
50:51you. I'm going to answer a real email
50:53right now without typing. You can see
50:55here Taiga is asking for feedback on our
50:57intro video for the next episode. Watch
50:59how I respond. If you could just make a
51:01few changes, uh make the red a little
51:03bit brighter,
51:04>> [music]
51:04>> the typography a little bit smaller. Uh
51:06for the music, if you can just give me
51:07some other options. And that [music]
51:09first photo, some more options as well.
51:11Uh please give me three options for each
51:13and uh make sure you send it to me by
51:15end of day Friday. Thanks.
51:17So, as you can tell, the text is way
51:19cleaner. [music]
51:20I kind of speak in half sentences and
51:22sometimes mumble on. This makes it very
51:24clean, formatted perfectly, and avoids
51:27all the small mistakes. This is a
51:28product that I cannot live without. I
51:31promise you that you will love it. If
51:33you guys want to get 1 month free of
51:34WhisperFlow Pro, click the link below
51:37and use code Open Residency.
51:41So,
51:42yeah, context. We've we've we've taken
51:43our really capable stranger that we've
51:45employed and we've
51:47to told him everything he needs to know
51:48about us.
51:49He's ready to start working for us, but
51:51he's still going to be pretty useless if
51:52you don't give him like tools to work
51:54with. Like imagine that you've just got
51:55this this dude that's working for you,
51:56but it's like he doesn't have an email
51:58account or a computer or browser or a
52:00phone. He's like still going to be
52:01useless. So, we need to give him tools
52:03to be able to actually complete the work
52:04for us.
52:06So, the way we do this is what's called
52:08a a thing called MCP. That stands for
52:10model context protocol. And I I don't
52:12want to lose anyone here. It's it sounds
52:14technical, but
52:15it's it's super simple. And and this all
52:17MCP is is just a way for your agent to
52:20connect to tools.
52:21>> The third-party tools.
52:23>> Yeah, the third-party tools. Yeah,
52:24exactly. Things like, you know, Slack,
52:26Gmail, calendar. Like you need a way for
52:28your agent to to connect and you do that
52:30through what's called MCP.
52:32>> To give people more context as well, I
52:33think a lot of people out there use
52:34either N8 or Zapier is a big one that
52:37people use to kind of extract
52:38information and and connect it.
52:40>> Yeah.
52:40>> And we're in the era now. I'm so curious
52:42to see. I think all the principles here
52:44are evergreen.
52:45>> Yeah.
52:45>> But it's like it's a big deal now to
52:47have an MCP. And like some of these big
52:48companies don't have MCPs.
52:51And like it's it's They're crazy. We're
52:53going to get into the BI of MCP.
52:55>> It It's what I make purchasing decisions
52:56on now. If if I'm choosing Let's say I'm
52:58choosing like a a marketing email
52:59marketing software, for example. If like
53:02if I've got five options and if if they
53:04don't have an MCP for a way for me to
53:07connect my agent to it, I
53:08it's out of the question. I won't use
53:10it. I don't use front ends anymore of
53:12apps. Yeah. Like like the future of this
53:16is And everyone that I know that that
53:17works the way this new way of working,
53:19this stage two, which is agents, they
53:21don't really use like front ends
53:23anymore. Like I complete a full work day
53:25from within inside like my cloud code or
53:27or agent harness that I'm using.
53:29>> By front ends, just so people out there
53:30listening, like you're not in platform,
53:32in Be Highve, on the dashboard.
53:34>> Yeah, exactly. You're extracting from
53:35there and then working within your
53:36workspace. Yep. Yep. Like I I use Google
53:38Calendar, I use Gmail, I use Slack, I
53:41use Stripe for payments, Calendly.com
53:42for bookings, and I can't even tell you
53:45the last time I've opened any of those
53:46apps up. They're just all connected into
53:48Claude. It's this new It's a It's a
53:49completely new way of working where
53:51you've got
53:52Claude or whatever AI harness you're
53:54using that sits on this as like an
53:56intelligence layer on top of all your
53:58tools, and then you're just interacting
54:00with like this
54:01intelligence layer.
54:02>> So, you know, I've alluded to MCPs being
54:05just this this way for your agent to
54:07connect to tools, but
54:08>> super quick on like what it actually is,
54:10right? So, before MCPs existed, if you
54:13wanted to connect your agent like Claude
54:15to a tool,
54:16it was it was difficult. It was
54:18possible, but required a lot of custom
54:19development. And the way to think about
54:20it is like Claude speaks English, right?
54:22And all these different tools speak
54:23different languages. Like Gmail speaks
54:25French, Notion speaks Spanish, Slack
54:27speaks Chinese. So, if you want Claude
54:28to talk to to Gmail to like get it to do
54:31things,
54:33it was required a lot of like custom
54:34development. It was It required time and
54:36effort because Claude had to
54:37>> Heavy, slow, expensive, everything.
54:38>> French. Yep.
54:39>> And shout out Ross Mike. This is
54:41explanations from him. He was on like
54:43some pod I heard him. So, this is taken
54:45directly from him. This is what landed
54:46like really nicely in my brain. But MCPs
54:48are just like this translator that sits
54:50in the middle between Claude and your
54:52and your tools. So, Claude does not need
54:54to learn French and Spanish and Chinese.
54:56It just has this little translator in
54:58the middle, so it just makes calls and
55:00then it can be relayed really really
55:02nicely to the tools. This is
55:03oversimplification, but quite frankly,
55:05like no one really needs to know what
55:07happens here. All you need to understand
55:08is that MCPs are just your way to
55:10connect tools into your agent.
55:12>> Uniformity, simplicity, and it just
55:14streamlines everything for
55:15>> the speed as well.
55:16>> Same way you can think of it like an
55:17API, right? For those of you who know
55:19what an API is. Like it's just a a way
55:21for tools to talk to each other.
55:23>> And again, that's an oversimplification
55:24of explaining MCPs and connectors. And I
55:27think the big thing to take out of this
55:28episode is the ability that you can ask
55:30the LLM like, "Oh, how do I set up
55:33connectors and how I do it?" And I don't
55:35want to encourage people to be lazy, but
55:36you want to just quickly just show over
55:38the screen of just where you go into
55:39Cloud to do it because it's very very
55:41simple.
55:41>> It's so simple. So, the way like if if
55:43we open the Cloud desktop app
55:45and we go into customize, you can see
55:47straight away connectors.
55:48>> Oh.
55:49>> And we have like if we go up into here
55:51browse,
55:52there is so many connectors here. Like
55:55pretty much all of the major tools that
55:57you would use have a connection here.
55:59And it's as simple as this, right? So,
56:01let's just say we want to connect up
56:03ClickUp. So, I can look You just go here
56:05and go plus
56:06and then I'll just one-click log in to
56:09my ClickUp.
56:12>> Guys, for context, we are power users of
56:14ClickUp. That is the harness, the
56:15project management tool
56:17that we use that creates all of the
56:19workflows, all of the dependencies.
56:22>> So, we can go in here. Literally, I've
56:24selected my workspace, connect.
56:26Open Cloud, it's connected. We go back
56:29to the desktop, it's connected to
56:29ClickUp. And you can see here like we've
56:31got all of these different things that
56:33we can do in ClickUp, create tasks,
56:35create lists. Cloud can now pull
56:37information from ClickUp if you're
56:38planning like marketing campaigns, it
56:39can update it. And we've got all these
56:42like by default they're set to needs
56:43approval, so it's always going to ask
56:44you before it takes actions. But as you
56:46start to build up trust with Cloud and
56:48like
56:48>> Yeah.
56:48>> you know, you you get more comfortable
56:50with allowing it to have more access,
56:51you can just do always allow. Like all
56:53my tools I just have always allow pretty
56:54much. But in Cloud, right? That's as
56:57simple as it is to connect these. But
56:59that is using MCP, by the way. So, you
57:01don't really have to deal with any of
57:03the technical [ __ ] But that when you
57:05click the plus and log in, that is it's
57:07it's got an MCP server, basically.
57:09That's what it's using.
57:10>> It's crazy that you say that like you're
57:11picking vendors based on if they have an
57:13MCP or not. It's literally like five to
57:1510x. We're going to dive deeper on a
57:17specific use case onto how much more
57:19granular and strategic and faster and
57:21optimized it is with this.
57:23>> Yeah, I mean
57:24it begs the question of just like
57:27are a couple LLMs going to win and then
57:29just the world is just going to be a ton
57:30of MCPs and connectors into it?
57:32>> Yeah, like if you're building a software
57:33product, it's like I I had a I had a
57:35buddy come to me and he's like, "Yo, I'm
57:37building this like travel agent like
57:39AI agent that helps me like plan their
57:41trips." And he's got this like nice
57:43website where people log on and you you
57:45say, "Okay, I'm going to New York for 10
57:46days. Like help me do this." And it like
57:48has an agent that plans your trip.
57:51Like I I said to him, I'm like, "I I
57:53don't think that's the right way you
57:54should be building products for the
57:55future." I'm like, "You'd be better off
57:57building this travel agent as an MCP
58:01that you then get hooked up into Claude.
58:03And then now people who use Claude can
58:06now have you show up as a connector
58:08here. Click plus and with your, you
58:10know, travel AI and then add your
58:13product as an MCP. And then now you
58:15build a purpose-built MCP for Claude
58:17that helps that makes Claude be able to
58:20plan your trips better." That helps
58:22Claude like gives it some extra tools it
58:24needs. Like it can search, you know,
58:25like Skyscanner and some other websites
58:27a bit easier.
58:28>> Yeah, I mean I feel like Yeah, I feel
58:30like the future is like as opposed to
58:31offering like a hundred things, it's
58:32like be really really dialed on 10, 20
58:35things. Then it becomes almost like an
58:36external data play in conjunction with
58:39that vertical.
58:40>> Yeah.
58:40>> Like
58:41be really, really smart and understand
58:43and know less things and just feed it
58:45into Claude. I completely agree. We talk
58:47all the time about like mid-market B2B
58:49SaaS is absolutely smashed. Like some of
58:51these companies that got sold, like I
58:53don't even want to call some of them
58:54out, but like basically what I was doing
58:56on these SaaS tools,
58:59I'm literally just doing it inside
59:01Claude right now.
59:02>> Yeah. Yeah, exactly. I mean like
59:05Yeah, the the the SaaS market is just
59:07getting absolutely decimated. And I
59:08think another thing is like a lot of
59:09these softwares build on like head
59:11count. Like they build on like seats in
59:13the software and now it's like they're
59:14just people like reducing head count.
59:16They got agents in there and like they
59:17can't bill for like agents. It's
59:20it's bad.
59:20>> Yeah, so you mentioned kind of the
59:23always allow or kind of ask first. Like
59:25that begs to ask the question of like
59:27security. A lot of people are very
59:28concerned about like connecting a lot of
59:30this stuff. Like what's your overall
59:32thesis on how people should approach
59:33security?
59:35>> You know,
59:37you just have to weigh up the trade-off.
59:39And for me, it's like yeah, there's a
59:41risk of things going wrong at like maybe
59:44sending an email to someone who it
59:45shouldn't without my approval. Of
59:47course, there's a risk, but it's like
59:50I've I've weighed up carefully like that
59:52risk compared to these productivity
59:53gains that I'm making. And to me, it's
59:55worth it. And I guess you just have to
59:56do that in your own head. But like
59:58there's of course there's so many edge
59:59cases with this. It's like you hear like
1:00:01horror stories about agents doing things
1:00:02for people. And it's like
1:00:04you hear the same thing about people
1:00:05getting hacked on the internet. It's
1:00:06like just cuz someone got hacked or
1:00:08you've been hacked before, doesn't mean
1:00:09you're not like going to stop using the
1:00:10internet, you know?
1:00:11>> Yeah.
1:00:12>> Definitely edge cases. And I think like
1:00:13the last line is like data is one thing.
1:00:16I think like connecting your something
1:00:18that has like your credit card and it
1:00:19can like go buy things. I feel like that
1:00:21is kind of the last line of defense. But
1:00:23like
1:00:24>> You could just do like for a bunch of
1:00:25things like draft only or take it to the
1:00:27finish line and don't click the button
1:00:29at the end.
1:00:30>> Yeah, so I would say like if you're
1:00:31really like conscious about it, like
1:00:33give it these tools, but give it what's
1:00:35called a read-only access. So it can't
1:00:37do actions, it can just read the data.
1:00:39For example, read emails, but it can't
1:00:40send anything. And that's just a nice
1:00:42way to start out and then you can like
1:00:44increase the permissions as you start to
1:00:45like feel more trustworthy. I give
1:00:48Claude full access to everything.
1:00:49Stripe, everything. It's It's got full
1:00:50access. I'm comfortable doing that and
1:00:52I've It's a risk I'm willing to take on.
1:00:54I've been using it for like 8 months and
1:00:56I've never had any issues. Some of those
1:00:58other harnesses like Open Claude and
1:01:00like Hermes and stuff, they are built in
1:01:03a way where the agents like not going to
1:01:05ask you for permission as much. The way
1:01:07that like these big companies like
1:01:09Anthropic and and and OpenAI have built
1:01:10like Codex and Claude code, the models
1:01:13are very good at asking you for
1:01:14permission for stuff. Like it pretty
1:01:15much anytime. Sometimes it gets annoying
1:01:17for me. It's like I'm trying to get it
1:01:18to answer emails, but it'll ask me for
1:01:20permission. Hey, "Hey, I send this? I'm
1:01:21like, yeah, send it." Like,
1:01:22I've given you the access.
1:01:23>> to like allow access, allow access,
1:01:24allow access all the time. Yeah.
1:01:26>> Yeah, yeah, yeah, yeah. Exactly. So, I
1:01:28I've never had any issues personally,
1:01:29but it's just something you have to
1:01:30weigh up. I mean, like there's two sides
1:01:32with security questions. It's like,
1:01:34obviously,
1:01:35giving it your tools and then it going
1:01:37and like taking actions. That's That's
1:01:39probably my least concern like my my
1:01:41least concern is it taking actions on my
1:01:43tools cuz that's like not that big of a
1:01:45deal. The other security side of things
1:01:47is like the models and like putting
1:01:48sensitive data through it. This is like
1:01:50a big thing and like a big problem I'm
1:01:52trying to like solve with working with
1:01:53some of these large like enterprise
1:01:54companies.
1:01:55Is it cuz some of them have like
1:01:56literally said like we don't want any
1:01:58personal information, company IP running
1:02:00through a Claude model cuz they're not
1:02:02convinced that like the Claude isn't
1:02:04going to steal their data. Yeah. So,
1:02:06that I mean, that's a whole whole
1:02:08another question, but for like the
1:02:09personal level, just like getting the
1:02:10tools connected and getting it to start
1:02:12doing things for you,
1:02:13yeah, just
1:02:14start with read-only access if you don't
1:02:16feel comfortable and then you can
1:02:18escalate it up as you go.
1:02:19>> Nice.
1:02:20Um let's talk about a couple of the
1:02:21tools that are like must-haves for you
1:02:24on your side. I want to share some of
1:02:25the stuff on our side as well, too.
1:02:26>> 100%.
1:02:27So, if you think back to like our little
1:02:29agent loop of observe, think, act, we've
1:02:31tuned up that
1:02:33observe step with context and now we're
1:02:34going to tune up the act step, right?
1:02:37These are like the top six things I
1:02:39reckon that everyone should have
1:02:40connected up to their agent or Claude,
1:02:42whatever you're using.
1:02:44First one is Apify. So, Apify is
1:02:47essentially a library of online
1:02:48scrapers. And these are like scrapers
1:02:50that people have built. Let me just go
1:02:53into here.
1:02:54>> So, it's almost like a third-party like
1:02:55sub-agent that helps you scraping.
1:02:57>> Yeah, exactly. So, it's like a
1:02:58marketplace. There's thousands of
1:03:00different scrapers you can use here. So,
1:03:01you can see on Apify, right? We've got
1:03:04Instagram profile scraper, YouTube
1:03:05transcript scraper, Reddit post scraper.
1:03:08>> Wow. This eliminates like some people
1:03:09are using like Grok as an LLM under the
1:03:12notion that it could scrape Twitter/X,
1:03:14but now How can just use Apify.
1:03:16>> Exactly. And we've got LinkedIn post
1:03:18scraper, Google Maps scraper, there's
1:03:20like Craigslist scrapers. Uh you can get
1:03:23like you can get lead scrapers, email
1:03:24scrapers.
1:03:25>> all individual MTPs?
1:03:27>> No.
1:03:28>> Oh, it's just Zapify.
1:03:29>> Zapify. And
1:03:30>> Damn. Zapify cut the chat. [ __ ] I never
1:03:33heard of this one. Is it good?
1:03:34>> Dude, it is everyone needs it. It like
1:03:36you literally there's there's there's
1:03:37almost nothing that Claude like can't
1:03:39scrape off the internet for you with
1:03:40Zapify connected.
1:03:41>> Damn.
1:03:42>> This is one that no matter what tools
1:03:43you use, everyone I think needs this.
1:03:45They've got a pretty generous free plan.
1:03:47It's like a
1:03:48five Yeah, like $5 a month worth of
1:03:50usage.
1:03:52>> Yeah, we're going to throw another zero
1:03:53or two behind that. This is crazy. Holy
1:03:55[ __ ]
1:03:56>> So, Zapify, everyone needs to have that
1:03:58one set up.
1:04:00The next one is Fire Crawl. So, by
1:04:01default, like pretty much all of these
1:04:03agent harnesses as part of the act step,
1:04:05one of the tools they all have baked in
1:04:06is a web search. But, this is super like
1:04:08simple and primitive. It's like if you
1:04:10were to ask the agent, like what's the
1:04:11price of Bitcoin right now? It can just
1:04:13web search and get the price. It's like
1:04:14more information gathering.
1:04:16When it comes to like things like hey,
1:04:19scrape the branding for this website, or
1:04:21like go look at this website and tell me
1:04:24like why they're or like do a CRO audit
1:04:27of my website.
1:04:28>> Less objective, more subjective, more
1:04:30like contact strategy type stuff. Not
1:04:32like what's the This is the question,
1:04:33this is the answer. It's more
1:04:35>> No, no. I I would actually say it's it's
1:04:36more that like the web search, it can't
1:04:38look at the website. It can't actually
1:04:40like Does that make sense?
1:04:42>> Web search can't extract raw data
1:04:44almost.
1:04:45>> Yeah, yeah, exactly. Like web search is
1:04:46just to like find like facts off the
1:04:48internet.
1:04:48>> Yeah.
1:04:49>> Whereas Fire Crawl allows it allows it
1:04:52to actually like read the page, not just
1:04:54search it. Like look at the page. So, it
1:04:56can extract like it can tell what the
1:04:58branding colors are. You could say like
1:05:00do a CRO audit of the page, and then it
1:05:02can use Fire Crawl
1:05:04>> That's the first thing my brain went to
1:05:05from a CRO and merchandising perspective
1:05:06from an e-com store.
1:05:07>> Yeah.
1:05:08Yeah.
1:05:08>> Yeah, exactly. So, you know, if you if
1:05:09you said if if you if you didn't give
1:05:11Claude Fire Crawl, and you just had
1:05:13default web search and you're like yo,
1:05:15here's my e-com store like run a sale or
1:05:17order. It can't. It can't see the store.
1:05:19It can just like maybe like search some
1:05:21information about your website.
1:05:23Like what you sell or like a little
1:05:25about section.
1:05:27With with with firecrawl it can actually
1:05:29like look at your page, scrape the
1:05:31entire thing. It can see like what
1:05:32buttons you have and where and like you
1:05:34know, it could see like your your
1:05:36buttons are the same color as the
1:05:37background like first fix make the
1:05:39buttons like brighter. So like stuff
1:05:41like that. It can give Claude like the
1:05:42ability to actually like scrape websites
1:05:44properly.
1:05:46And across across the default web
1:05:47search, firecrawl, and Appify there's
1:05:48like pretty much nothing that Claude
1:05:50like can't
1:05:51do for me across the internet.
1:05:52>> And this is just mitigating the back and
1:05:54forth of just bottlenecks. It's just
1:05:55getting you more and more and more
1:05:56context.
1:05:57>> Yeah. Yeah. And it's less it's less um
1:05:59human in the loop steps, right? Like if
1:06:00you didn't have Appify or firecrawl
1:06:02connected, it used Appify before when we
1:06:04did the scrape my Instagram account
1:06:06example.
1:06:07>> Damn.
1:06:07>> So like if I didn't have Appify
1:06:09connected, it
1:06:10>> probably would have said like yeah, I
1:06:11can't connect to Instagram like can you
1:06:13send me the videos or something.
1:06:14>> Context for people out there listening
1:06:15too like you don't have to say anything.
1:06:16This is embedded in the brain on the LLM
1:06:19platform.
1:06:20>> Yeah. Yeah.
1:06:21>> Damn.
1:06:21>> You can think of like you can think of
1:06:23our really capable employee that we've
1:06:26just hired. You can think of him as he's
1:06:27got his desk, right?
1:06:29>> Yeah.
1:06:29>> And
1:06:30he's got next to him he's got a
1:06:32bookshelf full of skills which we're
1:06:33going to get to. But he's also like got
1:06:35the tools like laying around like he
1:06:37might have like a you know, he's got his
1:06:38computer. He can see like he's almost
1:06:41got this up, right? We got applications.
1:06:45You can imagine that our ultra capable
1:06:46employee has this screen up, right? So
1:06:49anytime we give him a task, he doesn't
1:06:50you don't need to tell him go use
1:06:51Appify. It's like if you tell him go
1:06:53scrape this brand's Instagram page and
1:06:56give me the top best performing videos.
1:06:58>> He's going to just pick the right
1:06:58website.
1:06:59>> just going to look at his apps. Okay,
1:07:00cool. Like I should probably use Appify
1:07:02and then he's just going to go use it.
1:07:03So you don't need to actually specify
1:07:06The next one is Composio. Now, this is
1:07:08like some absolute sauce I'm going to
1:07:10put you on, right?
1:07:11Some quick backstory about how I found
1:07:13out about this wonderful tool. I had
1:07:15everything like baked into Claude. I had
1:07:17all my connectors built into Claude like
1:07:19I showed you before. And Claude went
1:07:21down for an hour the other week like the
1:07:23servers were down. And I had like so
1:07:24much work to do. I'm like, "Oh my god. I
1:07:26I And I And I can't work without Claude
1:07:28now."
1:07:29So, I had to open up like ChatGPT's
1:07:31Codex to like work in there. But
1:07:34And And the way that my folder is set
1:07:36up, you know, we we went back to the
1:07:37demo folder before. I've got like this
1:07:39OS folder. It stands for operating
1:07:41system where I work out of which I'm
1:07:42going to
1:07:43show you guys after this.
1:07:45So, the way that that's built is I own
1:07:47all the context files. I own all the
1:07:49skills. They're not embedded in Claude.
1:07:51I can open that folder up in Codex and
1:07:53keep working, right? That That part was
1:07:55fine. The issue that I found was I had
1:07:57like 50 different tools connected to
1:07:59Claude.
1:08:00And then I started to get Claude to
1:08:02Codex to work and it like didn't have
1:08:04any of my tools.
1:08:05So, I started to go through and add in
1:08:07like all these tools and and I was like
1:08:09"There must be a better way." That's how
1:08:10I came across Composio. So, what
1:08:12Composio does is we can open this up and
1:08:14you can see connect apps.
1:08:16They have way more than than the Claude
1:08:19desktop app. They have
1:08:20thousands of tools that you can connect.
1:08:23Same way with Claude like I showed you,
1:08:24it's one-click login. You click connect,
1:08:26sign in. They've got everything under
1:08:28the sun here.
1:08:29>> And then there's this like an MCP for
1:08:30MCPs?
1:08:31>> Yeah, it's like an Yeah, it's like an
1:08:33MCP for It's like a bridge, basically.
1:08:34>> Yeah. Yeah. So,
1:08:37I will connect all my tools in. Another
1:08:39thing Another thing is in the Claude
1:08:40desktop app, you can connect Gmail,
1:08:42right? And log in and you've got Gmail
1:08:44connected. But if you've got three Gmail
1:08:46accounts, you can't
1:08:47connect all three in Claude.
1:08:49Pain in the ass.
1:08:50And which most people have like multiple
1:08:51email accounts. So, in here you can see
1:08:53I've got three active Gmail connections.
1:08:54That was another thing that I liked
1:08:55about it.
1:08:56>> So, in short, you're basically saying
1:08:57that connect all of your MCPs to
1:08:59Composio and then if you're in a
1:09:01Composio
1:09:02>> then you yeah And then Composio is like
1:09:04that single MCP with all the MCPs
1:09:06underneath it.
1:09:06>> Damn.
1:09:07>> So, you can see here in in Claude in
1:09:09connectors Composio. This one won't show
1:09:11up as a browse connector. You'll have to
1:09:12click add custom and then add the URL,
1:09:14which I'll show you how to do later.
1:09:17And
1:09:19Yeah, we can see I've just got Composio.
1:09:20So, I could like theoretically just have
1:09:22that. I don't need the others. You can
1:09:24even connect the other ones I'm talking
1:09:25about Remember I'm talking about Appify
1:09:26Fire Crawl? You connect it through
1:09:27Composio.
1:09:29So
1:09:30>> Damn.
1:09:32>> Composio, that means like for my setup
1:09:35trying to like be, you know, robust and
1:09:37protect myself against like platforms
1:09:39going down.
1:09:39>> to go to your bunker in Argentina cuz
1:09:41the world is crumbling, you can just you
1:09:43can just hit up Composio and just bring
1:09:44all your [ __ ]
1:09:45>> Exactly.
1:09:46>> The MD files are skills of Composio.
1:09:48>> [ __ ]
1:09:48>> So, when I in future now if I want to
1:09:51work in Codex, I don't have to connect
1:09:52all these tools again. I just have
1:09:53Composio and then
1:09:55if I want to try out Let's just say like
1:09:56Flexity Computers is another one. I
1:09:58haven't really used it much. It's just
1:09:59another card.
1:10:00Different flavor of the same thing. But
1:10:02if I decided I wanted to use that, I
1:10:04could just I'll just connect Composio to
1:10:05that one time. Yeah, and then it's
1:10:07everything's hooked in. So
1:10:09>> For everybody out there listening, I
1:10:10want to really really stop and think.
1:10:13Like, where my brain used to go This is
1:10:15the problem that I used to have with my
1:10:16team. It's like, okay.
1:10:18I've never been on Codex. So, if Claude
1:10:21went down and I have to go elsewhere,
1:10:22like I don't even know if it's as simple
1:10:24as just clicking the connector to get
1:10:26Composio, but you could literally just
1:10:27go into Chat GPT and just say how do I
1:10:30connect Composio. And I think everyone
1:10:32just needs to understand that all the
1:10:35answers
1:10:36all these in between the lines answers
1:10:38are literally in the LLMs. You just tell
1:10:39them this is what I need done and they
1:10:41can do it for you.
1:10:42>> That's how I learned all this.
1:10:43>> Yeah.
1:10:43>> Like like the only like all this stuff
1:10:45I'm sharing with you today I just it's
1:10:46it's pretty much all learned from asking
1:10:48Claude.
1:10:48>> Yeah, you people need to go through the
1:10:50exercise of like if you're not a
1:10:52technical person and you'd have to bug
1:10:54somebody on your team to do something
1:10:55small for you
1:10:57force yourself to not bug them, and what
1:10:59may take you or them a minute, 2
1:11:02minutes,
1:11:03>> Yeah.
1:11:03>> it'll take you 5 minutes, 7 minutes, 10
1:11:05minutes if you're not that technical,
1:11:06but you will get to the answer. So, it's
1:11:08more just like
1:11:09>> Yeah.
1:11:09>> figuring this [ __ ] out by yourself is
1:11:11huge.
1:11:12>> This is like yeah, massive unlock to get
1:11:13in the habit of just asking Claude any
1:11:15of this stuff. It's like, "How do I do
1:11:16this?" Like, just just ask Claude. It'll
1:11:18It'll tell you.
1:11:19>> 100%
1:11:20>> And then if you start following the
1:11:20steps and get stuck, and you just
1:11:22screenshot this error, paste it in
1:11:24Claude, yeah. Hit this, like what do I
1:11:26do? Like
1:11:26>> That is a What you just said right there
1:11:28is a bar on its own. That's a cheat code
1:11:30on its own. It's just doing that. It's
1:11:31just screenshot, this is where I'm at,
1:11:33can't get past it, what do I do?
1:11:34>> Yeah.
1:11:34>> And let's keep going. So, these Chrome
1:11:36DevTools.
1:11:37>> So, the next one's Chrome DevTools. So,
1:11:39this is a an MCP built by Google Chrome.
1:11:42Maybe we can connect that through
1:11:43Composer. I don't know, I haven't tried
1:11:44it. But, essentially what this does is
1:11:45it allows Claude to spin up its own
1:11:48browser, if that makes sense. So, like I
1:11:50use this a lot if you're building like
1:11:51websites or apps.
1:11:53And Claude like wants to check its own
1:11:54work. Cuz you know, you might say it's
1:11:56like build a website, you know, these
1:11:58colors make it look nice, here's the
1:11:59brand guidelines. It might build it and
1:12:01send it to you, and it's just like
1:12:03wrong. It's not at all like what you
1:12:05wanted.
1:12:06Chrome DevTools is like one that I use
1:12:08that allows Claude to like be able to
1:12:09check its own work in a sense. Like it
1:12:10can spin up a browser and look at the
1:12:12website it's building and go, "Ooh, no,
1:12:13that's not good. I'll keep working on
1:12:15this."
1:12:15>> to like in-platform or a PDF file, it's
1:12:17like an actual browser.
1:12:18>> Yeah, it's an actual browser. And like
1:12:20another thing that I use it for is I
1:12:22don't know, another example earlier I
1:12:23saw on X someone was printing out like a
1:12:25register for like founders in Australia.
1:12:27Like almost like a phone book of of
1:12:29founders in Oz.
1:12:30And I clicked I was like, "Oh, this is
1:12:33cool. I'll sign up." And then it was
1:12:34like a you know, bunch of questions, and
1:12:37it was going to take me like probably 20
1:12:38minutes to sit there and answer them
1:12:39all. So, I just like got the link,
1:12:41pasted it into Claude. I'm like, "Yo,
1:12:43can you apply to this website for me?"
1:12:45And I just kept doing other work. And
1:12:46then because it has all the context
1:12:48about me, it knew exactly how to answer
1:12:49all the questions perfectly and in my
1:12:50voice. And but like without Chrome
1:12:53DevTools, it couldn't have had a browser
1:12:54to like go and actually enter it in.
1:12:56Little bit different to Firecrawl.
1:12:57Firecrawl's like more scraping, but this
1:12:59is actually like it can take actions on
1:13:00a browser.
1:13:01So, Chrome DevTools is a great one to
1:13:03get set up. These are pretty much all
1:13:05free, by the way.
1:13:06Another one is Playwright. This is one I
1:13:08use. Same vibe as Chrome DevTools. It's
1:13:10just another browser for Claude to use.
1:13:12Chrome DevTools spins up a fresh browser
1:13:13so it doesn't have any of your sites
1:13:15logged in, right? So, if you ask Claude
1:13:17to go send 1,000 LinkedIn connections to
1:13:18this list,
1:13:20and it's just got Chrome DevTools, it'll
1:13:22pop up the browser and it'll be like
1:13:23LinkedIn login and it'll just get stuck.
1:13:25Whereas
1:13:26>> And then you know, you might log in on
1:13:28that browser. Yeah. But then in a week's
1:13:30time, if you go to the task, it spins up
1:13:31a new browser again. It doesn't have the
1:13:33login saved. So, Playwright just allows
1:13:35it to control your current browser.
1:13:37And that like hypothetically, like that
1:13:39could automate like one password with
1:13:41filling in the password's advanced
1:13:43fields for like your address, where you
1:13:45want to ship it to, stuff like that.
1:13:46>> On on one password real quick, that's
1:13:47another one. I've connected one password
1:13:49to Claude. So,
1:13:49>> Risky business, bro.
1:13:51>> Yeah.
1:13:51>> That's a risky one.
1:13:52>> Yeah, it is.
1:13:53>> I don't recommend that for people out
1:13:54there. That's That's like the last
1:13:55That's the last level.
1:13:56>> Yeah, what happens though when it's on
1:13:57my computer and it's like getting a
1:13:59password, it pops up with the Touch ID
1:14:02and I have to like approve it with my
1:14:03finger. So,
1:14:05>> I stand corrected. I like that.
1:14:06>> Yeah. Yeah, that Yeah, that's kind of
1:14:07cool. I feel like Iron Man or something.
1:14:09>> So, on my computer, I just tap my
1:14:10finger.
1:14:10>> [laughter]
1:14:11>> That's a different story. Higgsfields
1:14:12the chosen one.
1:14:12>> Higgsfields the last one. So,
1:14:14Higgsfields are just like
1:14:15Dude, they're they're dominating.
1:14:16They're crushing the AI content space.
1:14:19Like they're like
1:14:21Their playbook is just like I don't
1:14:22know, needs to be studied. Like they've
1:14:24just done an incredible job with like
1:14:26influencer marketing and like just
1:14:28they've made themselves at the name as
1:14:30the top place to go for like AI content,
1:14:33like images and video.
1:14:34>> Yeah.
1:14:35>> So, you know,
1:14:36>> And that at its simplest form, guys, is
1:14:38you just prompt it and it creates images
1:14:39and videos for you. Like at at its
1:14:41simplest form.
1:14:41>> So, that's a great tool in itself. Like
1:14:43I have some buddies that like live in
1:14:45there and make some like insane AI
1:14:47content and like UGC videos and static
1:14:50ads and stuff. But if I default Claude
1:14:52does not have a way to generate images.
1:14:54So like chat GPT does have an image
1:14:56model whereas Claude doesn't. So this is
1:14:59especially one that's useful for Claude.
1:15:01I mean even chat GPT doesn't have video
1:15:02generation whereas Higgsfield is
1:15:04basically like the Composure but for
1:15:07image models. So it basically pulls in
1:15:09Google's top models, chat GPT's top
1:15:11models, like all the Chinese ones that
1:15:13are getting cracked as like it pulls
1:15:15them all in and allows you to just
1:15:16access them through one MC pair with
1:15:17Claude. So to show you the Higgsfield MC
1:15:19pair in action. Now this is just a demo.
1:15:21I don't I don't work with Bloom at all.
1:15:22Shoutout Greg. He's
1:15:23>> Shoutout to Uncle Greg. Great episode. I
1:15:24don't know if it's going to be it should
1:15:25probably be the episode before this.
1:15:27Absolute legend.
1:15:28>> Yeah. I had the pleasure of meeting him.
1:15:29He's a great dude.
1:15:30>> Great.
1:15:31>> So
1:15:32Bloom creates the gummies. I just gave
1:15:33it a product image and I was like yo,
1:15:35scrape competitors ads running in the
1:15:37meta ads library and generate 20 static
1:15:38image ads using the Higgsfield MC pair.
1:15:40And then it used Fire Crawl to scrape
1:15:42the meta ads library. And then it's
1:15:44connected to as well. This is like an
1:15:46official connector that meta has
1:15:48released so you can get your agents like
1:15:49through ads. Another great one I
1:15:50recommend but the Higgsfield one, right?
1:15:52You can see
1:15:54it's gone through and it started like
1:15:55generating static ads with Higgsfield.
1:15:58And like these are they're pretty good.
1:16:00I mean
1:16:02some of them
1:16:03it's always going to be the case with
1:16:04ads like it's going to you have to
1:16:06generate like a hundred and there'll be
1:16:07like maybe five to ten really good ones.
1:16:09So you can see here it's like it's done
1:16:10a bunch of different static ads and like
1:16:12the as a as an ads background I would
1:16:15run like that's a that's a that's a cool
1:16:17one. That mean that the the gummies
1:16:19don't look perfect but we could just
1:16:21have added an image of the actual
1:16:23gummies and then it would have nailed it
1:16:24perfectly. The like that's a great ad.
1:16:27The product's a little squished. That's
1:16:29a that's a [ __ ] phenomenal ad. That's
1:16:31a great ad too. Creatine for the girls.
1:16:33Like you can see here it's it can
1:16:34actually like use this MC pair to like
1:16:35start creating static ads. And um
1:16:38this is just off a really basic prompt
1:16:40but if I was trying to like automate as
1:16:43much of the ads process as possible,
1:16:45I'd build a bunch of skills around like
1:16:47static ads and like headline writing and
1:16:48copywriting and stuff, which we'll dive
1:16:50on soon. But, that's just to show you
1:16:51like the Higgs Field MCP and like
1:16:54uh some some use cases.
1:16:55>> Love it. Guys, this episode is sponsored
1:16:57by Ketone-IQ.
1:16:59I often sit down in this chair for over
1:17:013 hours and stay super dialed with
1:17:04ketones. It's a completely different
1:17:06category of fuel for your brain, clean,
1:17:08[music] sustained, and no crash. I take
1:17:11a shot before recording, before deep
1:17:13work, and any session that I need to be
1:17:14sharp. Head to ketone.com/openresidency
1:17:18for 30% off your first subscription and
1:17:20a free gift with your second shipment.
1:17:23That's [music] keton.com/openresidency.
1:17:27Guys, this is something that I use every
1:17:29single day, non-negotiable. Give it a
1:17:31try.
1:17:33This episode is brought to you by
1:17:35Momentous. Most Americans today are
1:17:37walking around deficient in omega-3,
1:17:39vitamin D, and magnesium. Three of the
1:17:41most foundational things in energy,
1:17:43recovery, and sleep. And even if you're
1:17:45already taking supplements, the industry
1:17:47is so loosely regulated that you really
1:17:49don't know what you're taking. I know
1:17:50the team behind Momentous personally,
1:17:52and that's why I trust it. Every
1:17:54product, every batch, is third-party
1:17:56tested [music] for NSF Certified for
1:17:58Sport. So, what's on the label is
1:18:00exactly what's in the bottle. It's why
1:18:01over 200 pro sports teams trust them and
1:18:05why my [music] whole stack is theirs.
1:18:06Omega-3, vitamin D, magnesium, and
1:18:08creatine every single day. If you're
1:18:10putting real effort into your health,
1:18:12you want to know that it's actually
1:18:13working. Go to livemomentous.com, use
1:18:15code openresidency for up to 35% off
1:18:18your first order.
1:18:21>> All right, guys. So, we gave you guys
1:18:21those six core tools that you need and
1:18:23then obviously some MCPs that we both
1:18:25like individually and as a company.
1:18:27Let's transition into skills that's for
1:18:32the newbies out there. Just to retouch
1:18:33on it. It's basically SOPs for for AI.
1:18:36Yeah. Why don't you just dive into
1:18:39the playbook on how to utilize skills?
1:18:41>> All right, let's do it. So,
1:18:43I would say skills is probably like the
1:18:45one of the most important things to
1:18:47master with all these agents.
1:18:49So, where we're at now is we've got a
1:18:50like really capable new employee and
1:18:51we've, you know, given him all the
1:18:53context, he knows everything about us
1:18:54and the business. Now we've given him
1:18:55tools so he can do work. He's already
1:18:57now going to be great. You can now start
1:18:59giving him work, he's going to do an
1:19:00amazing job. But we want to be able to
1:19:02train him on our processes and the way
1:19:05that we do things.
1:19:07So, an example of skills, they're
1:19:09literally just you can think of them as
1:19:10SOPs for AI.
1:19:12The same way if you had a VA that was
1:19:14helping you with like customer support,
1:19:15for example, you would probably have a a
1:19:17Google Doc built out where it's like
1:19:19customer support triage
1:19:21workflow. Step one, log in to
1:19:23monday.com. Step two, navigate to this
1:19:26section in Monday. Step three, find
1:19:28tickets. Do, you know, you would outline
1:19:30it really really like like simply to a
1:19:32VA like so it's literally foolproof so
1:19:34they couldn't screw it up.
1:19:36We're literally just taking exactly that
1:19:38but giving it to an agent. And instead
1:19:39of a Google Doc it's just a markdown
1:19:41file.
1:19:42And to illustrate like the problem that
1:19:44skills solve and how useful skills can
1:19:46be, I'm just going to give a simple
1:19:47example.
1:19:49So, let's just say you need a proposal
1:19:51written and you've got your tools and
1:19:53and and context all set up and you start
1:19:55a new session and you're like, all
1:19:56right, create a proposal for this new
1:19:58client like here's the price, here's the
1:19:59offer, blah blah blah blah blah.
1:20:01Send it off.
1:20:02It will give you back like a PDF and
1:20:04like
1:20:05very rare it's going to one shot it
1:20:06exactly how you want. You're probably
1:20:07going to have to go back and say like,
1:20:08don't like how you put the price at the
1:20:10top, should be at the bottom instead,
1:20:11like or the logo looks a bit funky, like
1:20:13put it in the top left, like change the
1:20:15colors of the text from blue to black.
1:20:17>> This is the V1 world of chat, just back
1:20:19and forth, back and forth, back and
1:20:20>> Yeah, yeah, this is still like the Well,
1:20:21no, this is using agents still. You
1:20:22still got agents like I mean, this
1:20:24applies to chat too, yeah, but but I'm
1:20:26talking about like an agent here. Like
1:20:27you let's say you're in code or co-work
1:20:28or getting it to make this proposal.
1:20:31You have to just explain your
1:20:32preferences, right? And eventually,
1:20:34maybe after 15-20 minutes of back and
1:20:36forth, you're going to it's going to
1:20:37deliver this proposal that you're happy
1:20:38with, right? Great. You have the
1:20:40proposal written, but what happens in a
1:20:42week's time when you want to create
1:20:44another proposal for another client? And
1:20:47you
1:20:48start a new session and you're like, all
1:20:50right, create a proposal for this client
1:20:51now.
1:20:52It's it's not going to remember any of
1:20:53those preferences. I mean,
1:20:55you you could have it save it to that
1:20:57memory.md like we did earlier, but it's
1:21:00it's not the right place for it. You're
1:21:02you're you're you'd be you'd be making a
1:21:03de facto skill. The the best place to
1:21:06have that is in a skill.
1:21:07So,
1:21:08if you if you're starting a a fresh chat
1:21:10and said to create a proposal, it
1:21:12wouldn't like remember any of those
1:21:13preferences and you're just going to
1:21:14like not you have to go through that
1:21:16whole process of explaining everything
1:21:17over again, which is a pain in the ass.
1:21:19Whereas with skills, you would tell
1:21:22Claude, you know, save this proposal as
1:21:25like a skill. Save this process that
1:21:26we've just been through as a skill so I
1:21:28can use it again in the future.
1:21:29And Claude would build a proposal skill
1:21:32and it would save everything that that
1:21:34it just done. It would save like all
1:21:36your corrections. It would save like it
1:21:38would probably save the end result as a
1:21:40PDF into the skill as well as like a
1:21:42reference for it to look at. And then it
1:21:44packages it up nicely into into this
1:21:47like skill folder and saves it. So, now
1:21:50anytime in the future, if you ask it to
1:21:52create you a proposal, it's just going
1:21:53to like whip up this SOP that it's
1:21:55written for itself. Literally a markdown
1:21:57file of step-by-step guide and then it
1:21:59will just follow it to a T. And then you
1:22:01can basically just set like train
1:22:04your AI agent on exactly how you like
1:22:06things done and exact processes.
1:22:07>> So duplicatable and I think anybody
1:22:10that's advanced, I mean, how many skills
1:22:12do you have? I I would imagine an insane
1:22:14amount.
1:22:15>> Yeah.
1:22:15>> We have a lot.
1:22:16>> Yeah. Yeah, I've I've got a lot.
1:22:18>> As far as this goes, like a very very
1:22:19easy example and we're going to dive
1:22:20into the different ways in which you can
1:22:22execute against this. Like let's just
1:22:24think like
1:22:24>> together. we're
1:22:25>> a live skill, guys. So, like even just
1:22:27think like social media captions, you
1:22:28know, from a formatting perspective, you
1:22:30can say hook, body, CTA, tone of voice,
1:22:34you know, I like to not be, you know, so
1:22:36professional.
1:22:38Get the hashtags out of there. And then
1:22:40an example output you have as well, too,
1:22:42is just like yeah, these are, you know,
1:22:4410 different examples of what it looks
1:22:45like. And I think like the key thing to
1:22:47take out of this is is repeatability as
1:22:50well. Like you execute it and nail it
1:22:52one time
1:22:54and then you're done.
1:22:55>> Yeah, pretty much. And the way you can
1:22:56like think of of how our really capable
1:23:00employee uses skills is, you know, I was
1:23:03talking about 40 times his desk and his
1:23:04computer and see his apps and stuff. And
1:23:06he's got this bookshelf next to him of
1:23:07SOPs, essentially.
1:23:09And
1:23:10the anatomy of a skill, let me actually
1:23:12open one up. I'm going to pop one open
1:23:13and show you what it looks like.
1:23:15So, if I go into like my OS folder,
1:23:17which is like my folder that I do all my
1:23:19agentic work in,
1:23:21I have
1:23:22this hidden folder called dot Claude.
1:23:25Don't let this scare you. This is just
1:23:26like literally a Think of this is like
1:23:28Claude's settings within that project.
1:23:31We can just open this up.
1:23:33And we've got this folder called skills
1:23:34with all of my skills that live in
1:23:35there.
1:23:37And I can go into, for example, like
1:23:40YouTube titles is a skill I've built,
1:23:42right? And we can actually open this up
1:23:45as a folder.
1:23:46So, so this is the skill in its
1:23:47entirety. It's a folder.
1:23:49And we open that up and what's inside
1:23:50the folder? We've got a skill.md file.
1:23:54And this is literally remember like the
1:23:55MD is just a markdown file. It's
1:23:57literally
1:23:58just a
1:24:00straight markdown file with my YouTube
1:24:02title process. And if we look at the
1:24:04anatomy of this skill, every skill is
1:24:06kind of broken into three parts. You've
1:24:07got
1:24:09the name, description, and then
1:24:10contents.
1:24:12So, the name at the top is just YouTube
1:24:13titles. The description is generate
1:24:16YouTube title options for a video topic.
1:24:18Use when the user says generate titles,
1:24:20YouTube titles, or title options for my
1:24:22video.
1:24:23So, you can see like this is separated
1:24:24from the main content by those two
1:24:26little lines, right? And the way you can
1:24:28imagine this in your head is so you've
1:24:29got this bookshelf that your agent has a
1:24:31bookshelf next to it.
1:24:33And
1:24:34the books are all like stacked in,
1:24:35right? And on the spine of the book is
1:24:38just the name and description of all the
1:24:39skills.
1:24:40So, he's got like, you know, this full
1:24:42bookshelf. He would have like the
1:24:43YouTube titles with that description,
1:24:45maybe YouTube thumbnails along with my
1:24:46other 100 skills.
1:24:48And then when we ask
1:24:50Claude or our agent,
1:24:52"Yo, can you uh write some YouTube ti-
1:24:55titles for this video?" He's going to go
1:24:56look at his bookshelf and go,
1:24:59"Titles." Read the description, use when
1:25:01user says titles. That's probably the
1:25:02one, pull it out. Then he'll read it.
1:25:04And then he'll load that con- like all
1:25:06this content in.
1:25:08You got to think about that example
1:25:09where I was showing like the the bar
1:25:10chart, how it fills up with context. And
1:25:12it's already like got a little bit
1:25:13loaded in each session, which is your
1:25:14your Claude that I'm doing and the other
1:25:16files.
1:25:17And then alongside that, he's got
1:25:20basically loaded in all the name and
1:25:21descriptions of every skill that you've
1:25:23got, which is like probably hundreds.
1:25:25And that sits in his like taking up a
1:25:29little bit of that of that context bar
1:25:30of all times.
1:25:32But it's not like if he if he loaded in
1:25:34the contents of every hundreds of
1:25:35skills, that bar you wouldn't even be
1:25:37able to like give it a message in. It'd
1:25:38be like 100% full. So, the way that
1:25:40skills work, it's called like
1:25:41progressive disclosure without getting
1:25:43too nerdy. And it's basically like each
1:25:45of the name and descriptions just get
1:25:47loaded in at the start of a new session.
1:25:49And then when the task like requires
1:25:52like you know, YouTube titles, and he
1:25:54sees the YouTube titles book, then he'll
1:25:56open that up and then it will load the
1:25:57full skill content into the session
1:25:59window. Yeah. Does that make sense?
1:26:01>> Yeah, and we're going to walk through
1:26:02that so people can kind of see it in
1:26:03action. If if we look at this
1:26:06title skill,
1:26:07you can see that
1:26:09it's literally an SOP like you would
1:26:11write for VA.
1:26:12Hey, step one, look at this, get the
1:26:14topic.
1:26:16Find the video topic.
1:26:18You know, if the user gives it to you,
1:26:20extract it from the the folder or file
1:26:22name.
1:26:23Step two, competitor research. Use
1:26:25Appify. It's telling it to use that
1:26:27tool. Remember Appify?
1:26:29To scrape top performing videos on the
1:26:31same topic.
1:26:32Target channels. And it's giving some
1:26:34ones that I've I wanted to to do. What
1:26:36to extract? The video titles, the title
1:26:38length, the structure,
1:26:40use of numbers, capitalization patterns.
1:26:43And then it's telling it like how to use
1:26:45Appify. By By the way, I didn't write
1:26:46this skill. Your Claude will always
1:26:48build skills for you. But you can see
1:26:50like this is just to illustrate is
1:26:51literally just a document with
1:26:53step-by-step guides that you're giving
1:26:54to your employee.
1:26:55And then step four, generate the title
1:26:57options.
1:26:58Categories, curiosity gap, benefit
1:27:00first, contrarian number. And this is
1:27:02this skills is where you package your IP
1:27:04up into. Like by default, Claude is not
1:27:07going to write a good If you go a new
1:27:08session with Claude and you're like,
1:27:09write me YouTube titles, it's going to
1:27:10suck. It's like the titles are going to
1:27:12be like so bad, so generic.
1:27:14>> Yeah.
1:27:14>> You I know you've tried it.
1:27:16>> I mean, did. We've tried all of this
1:27:17stuff, too.
1:27:19>> And it's like
1:27:21you package like I've got a way that I
1:27:22like my titles written that's unique to
1:27:24me and how I think about titles. And I'm
1:27:27putting that IP into this skill,
1:27:29teaching it how to think.
1:27:30And in the same way like if you're let's
1:27:32say media buyer, if you tell Claude like
1:27:34you're can you handle my Facebook ads
1:27:37for me?
1:27:38It's it's it's not going to know
1:27:40>> Your unique sauce. I mean, yeah. You
1:27:42know, we talked about it before like
1:27:44obviously there's taste and judgment and
1:27:46creativity that goes into kind of all
1:27:47these MD files and skills like
1:27:50enterprise value is going to live within
1:27:53like what these skills and MD files are.
1:27:56That's such a good probing question to
1:27:58kind of
1:27:59see how deep someone is in in the space
1:28:02is like what does that look like? Yeah.
1:28:06Been looking under the hood on our side.
1:28:07It's
1:28:08[ __ ] out of control. I can only
1:28:10imagine what yours is.
1:28:11So let's just take a step back on the
1:28:13skills side of
1:28:16like what it actually looks like to make
1:28:19these skills, whether you're intentional
1:28:20about it or kind of your LLM kind of
1:28:23does it for you. What does that look
1:28:24like?
1:28:24>> Yeah, so let's let's run through that.
1:28:26For creating skills, in my mind there's
1:28:28two ways to go into it.
1:28:30There's goal first or process first. So,
1:28:34the goal first way is to literally say
1:28:35up front, "Build me a brand guideline
1:28:37skill."
1:28:38And then here's the brand guidelines PDF
1:28:40and it'll go and build it for you. By
1:28:41the way, Claude has and pretty much all
1:28:44of these agent harnesses
1:28:46should come, most of them do, with a
1:28:48skill creator skill built in. So, it's
1:28:50like skill-ception. You've literally got
1:28:52a skill
1:28:53that has a step-by-step guide on how to
1:28:55create more skills. And that's like not
1:28:58one you need to download. It comes
1:28:59default. I know it does with Claude and
1:29:01most of the others as well. So, if you
1:29:04tell Claude like, "Hey, build me a skill
1:29:06for brand guidelines." He's going to
1:29:07look at his bookshelf and see the skill
1:29:09creator skill and like create it and
1:29:10then know what to do.
1:29:11>> Insane. So, in order to make like a
1:29:13brand guideline skill, which is a great
1:29:14one to give your agents so they can have
1:29:16like a just a repeatable playbook on how
1:29:17you do branding for your, you know, lead
1:29:20magnets, PDFs, websites, whatever.
1:29:22Trying to build that the process first
1:29:23way doesn't make sense. And I'll just go
1:29:25straight into like what the process
1:29:26first way is. So, you've got goal first,
1:29:28which is, you know, just telling it up
1:29:29front, "Build this." The process first
1:29:31way is the way I build 90% of my skills.
1:29:34And this will be like I will start a new
1:29:36session with Claude and I will, for
1:29:38example, like I'm going to demo my ad
1:29:39scraping skill. So, to build this, what
1:29:41I did was I started a fresh session and
1:29:43I'm like, "Okay,
1:29:44here is the ads library link for
1:29:47XYZ competitor.
1:29:49Please
1:29:51use Firecrawl to scrape like their
1:29:54entire like ad library and download it."
1:29:57Send. And it went and did it. And I was
1:29:58like,
1:29:59"All right. And now please analyze every
1:30:01single one of these ads and then give me
1:30:03a breakdown of like these metrics, the
1:30:05ones I care about." Send.
1:30:06>> I think for people too, if you're like a
1:30:07macro perspective,
1:30:09top-down versus bottom-up is basically
1:30:11what this is right now. It's just how to
1:30:13how to how to look at it as well.
1:30:14>> Yeah, yeah, yeah. Top-down versus
1:30:14bottom-up. That's a great way to put it.
1:30:15>> Yeah, yeah. It's a It's a It's a good
1:30:16way to put it.
1:30:17>> And then I'm like, "Yo, Claude, go find
1:30:20all the different landing pages they're
1:30:21running ads to." And it like sent it
1:30:23off. It went and found like, "Yo,
1:30:24they're they're running it off these
1:30:26five pages." Okay, use fire crawl. I
1:30:29want you to take a screenshot of all the
1:30:30landing pages and build a a master doc
1:30:33breaking down each of the landing pages
1:30:34they're running ads to.
1:30:36And then anyway, I went through this
1:30:37entire process. And then
1:30:40I once that was done and I'm happy with
1:30:42it, like I've taken Claude on this
1:30:44almost like a journey about how I want
1:30:46this process done.
1:30:47Then I will go like,
1:30:49"I know I'm going to want to do this
1:30:50again in the future.
1:30:52I'll just tell Claude, "Okay,
1:30:53create a skill for this process that
1:30:55we've just done and call it ad scraper."
1:30:57And then it uses the skill creator skill
1:30:59and and and packages it up for me. And
1:31:00now I have this skill, which I'm going
1:31:01to show you right now. It's It's
1:31:02absolutely insane. Blow my mind. Let's
1:31:04do it.
1:31:05Cool. So, you can see I've got the ads
1:31:07analyst skill. And I'm going to do a
1:31:08little demo here. So, how this skill
1:31:09works is I just have to give it an ads
1:31:10library URL and it will go through like
1:31:13I said before and scrape all the ads
1:31:14landing pages and build me like a full
1:31:16report.
1:31:17>> Under the notion that you already have
1:31:18it connected in your LLM. Like what we
1:31:20talked about earlier with the MCPs. It's
1:31:21It's It's
1:31:22>> Yeah. Yeah. It's using like the some of
1:31:24those tools like Appify and Fire Crawl
1:31:26to scrape them.
1:31:26>> Cool.
1:31:27>> So, this is my buddy Ollie. I was going
1:31:29to scrape his.
1:31:30>> Shoutout on Ollie, baby. Let's go.
1:31:32>> Ollie's Ollie's a gun. He was one of the
1:31:33ones that got me into into content.
1:31:35>> He's an OG of the AI space.
1:31:37>> All right, cool. So, in Claude code, and
1:31:38I needed to make sure I'm using my OS
1:31:40folder, which is where I pretty much do
1:31:41all my operations out of because that's
1:31:43where my skills live.
1:31:45And if we go into here, we can go slash
1:31:48and it will bring up all the skills we
1:31:49have available.
1:31:50>> Hey.
1:31:51>> Ads analyst, ready? Orchestrate
1:31:52competitive ad research, extract ads,
1:31:55create a report. And we just click that,
1:31:57ads analyst. Literally, we just give it
1:31:59the ads library link. Bang. And we just
1:32:01send it off.
1:32:02Go.
1:32:04And now that is going to start cooking
1:32:06away for us.
1:32:07So, Mark, I wasn't going to make you
1:32:09wait another 20 minutes for this to to
1:32:11run through. So, I actually ran the
1:32:12exact same thing earlier before with the
1:32:14same as library URL as analyst skill.
1:32:17And you can see it finished up the
1:32:20analysis. And it basically gave us a
1:32:21master report, a bunch of things. And it
1:32:24deposited it into this active folder
1:32:26that I have. So, within my OS folder,
1:32:28which is literally like a folder on my
1:32:29desktop, it's got a bunch of like sub
1:32:31folders within it where I kind of just
1:32:32like keep things organized. And I like
1:32:34to have this little active folder here
1:32:37where I just kind of just miscellaneous
1:32:38tasks that aren't really related to
1:32:40anything specific. It just creates a new
1:32:42folder to work in. So, it's You can see
1:32:44here we've got this ad research one
1:32:45where it's just sort of deposited in
1:32:47there. And we can open that up here, and
1:32:49it's built us a few things. It's given
1:32:51us a master report in markdown format,
1:32:53which we can open.
1:32:55And
1:32:56read through. You can see it looks
1:32:57exactly like a a doc you'd see in
1:32:59ClickUp or Notion. And we can go through
1:33:01as well.
1:33:03So, we can just like
1:33:03>> Wow, there's a linked version that links
1:33:05out to the individual ads?
1:33:06>> Yeah.
1:33:07So,
1:33:08>> Jesus.
1:33:09>> And you can see these other reports it's
1:33:10created. Also, by the way, it's
1:33:11downloaded all of the assets. So, you
1:33:14can see here like pretty much every
1:33:16single video.
1:33:17>> Jesus.
1:33:18>> I wonder if like in the desktop app
1:33:19it'll pull them up. But we can show them
1:33:20in Finder, right? It's downloaded every
1:33:23single one of his ads.
1:33:24>> Oh my god.
1:33:25>> And it has done like a deep dive into
1:33:27them all. It's pulled out all the
1:33:29landing pages. And it's basically, if we
1:33:31can open this up, I'll open that up in
1:33:33Cloud in a sec, but it's done like the
1:33:34full analysis. And you can see it's
1:33:36given us like this final report as HTML
1:33:39format. So, we can now just shut this
1:33:40up. And I'll open this up so it's a bit
1:33:43bigger for you to see.
1:33:44So, you can see that it's given us this
1:33:46full ads strategy analysis. 54 active
1:33:49ads, 37 video, 14 image, the price
1:33:52point, the longest running ad, newest
1:33:55funnel 3 days ago.
1:33:56And you know, it's a competitor of mine,
1:33:58even though I always say good good
1:33:59buddy, but same thing in the AI space.
1:34:02And I can like now go through and scrape
1:34:04like see everything that he's doing.
1:34:05It's It's basically talked about like
1:34:07things to steal or beat, like what he's
1:34:09done well. So, reframe the category to
1:34:11justify the price. Like he's selling
1:34:12like a Notion template, but it's not a
1:34:14template, it's an operating system. So,
1:34:16he's It's It's picked that up and it's
1:34:17given me that to me as a great takeaway.
1:34:20Name your mechanism, context
1:34:21engineering, so the method feels
1:34:23ownable.
1:34:24You know, longevity
1:34:25>> Yeah, I think for context, I mean, this
1:34:27is a a long document that we can spend
1:34:28all day on it. It's just like you could
1:34:30just keep optimizing this single skill
1:34:33and just
1:34:34on the skill side, like you can format
1:34:36it however you like. And obviously we're
1:34:38going to go through the whole entire
1:34:39gamut, but ultimately you can
1:34:41autonomously schedule updates on
1:34:43competitors like
1:34:44>> Yeah. Yeah.
1:34:44>> You can do anything you want
1:34:46in relation to this.
1:34:47>> I mean, look, I don't like the way
1:34:48that's formatted, to be honest. So, I
1:34:49would probably go back to Claude. And
1:34:51it's a good to point to touch on about
1:34:52is these skills. It's like the first
1:34:53time you build them,
1:34:55the V1 of the skill is never going to be
1:34:57perfect. All most of my skills that I
1:34:58use are on like V5 to 10. So, this one
1:35:01here is like I would then say to Claude
1:35:04in the chat, I hate how you've made the
1:35:06landing pages so like big and full
1:35:09screen. Like I'd like them to be, you
1:35:12know, smaller and better fitting on the
1:35:13page. And I would say just update the
1:35:15skill, and it will go and update it for
1:35:17you. You don't have to do anything. And
1:35:18then next time you run it, skill's on
1:35:19V2, you're not going to have that issue.
1:35:21And then you just keep doing that until
1:35:23the skill like is eventually sanded down
1:35:25until it's perfect.
1:35:26>> Keep iterating.
1:35:27>> Yeah. Yeah, I mean, we can go on and on
1:35:29about kind of reverse engineering the
1:35:30skills that you've already built, but I
1:35:31want to show everybody here and live how
1:35:34to actually have an idea and build a
1:35:36skill. We have a
1:35:37mutual connection here with Beehiiv. We
1:35:38both love Beehiiv. So, why don't we just
1:35:41have you from the bottom up? Let's just
1:35:43Let's just build a skill in conjunction
1:35:45with Beehiiv.
1:35:45>> Yeah. All right, great. So, let's let's
1:35:48do a One of the ideas we had was like a
1:35:51a line skill.
1:35:52>> Yeah.
1:35:53>> So, it will basically write subject
1:35:54lines for my newsletter by looking in
1:35:56Beehiiv at what's performed the best
1:35:58over the last like is historically and
1:36:00then use that to like write the subject
1:36:02lines. So,
1:36:04first up, we need to connect Beehiiv
1:36:06into Claude. So, I'm just going to show
1:36:08you super quick. We're just going to to
1:36:09go to connectors and now I don't think
1:36:11that the Beehiiv MCP is
1:36:14No, it's not native into Claude. So,
1:36:17what we're going to do is we're just
1:36:18going to go to plus at custom connector
1:36:21and then just
1:36:22like close out of these advanced
1:36:24settings. It's just like nothing to
1:36:25worry about. I'm just going to name it
1:36:27Beehiiv.
1:36:30Whoop.
1:36:31And then you need [clears throat] the
1:36:33the server URL here. And look, if you
1:36:35have a tool you want to connect to, you
1:36:36can just ask Claude like, "Hey, can you
1:36:38set this up for me? Or can you help me?"
1:36:40And it will walk you through it. But in
1:36:41my case, I already know what I'm doing
1:36:42here. So, I'm just going to go into
1:36:43Beehiiv into my account. We're going to
1:36:45go down to MCP in settings and they've
1:36:47got this URL for us to use. And like
1:36:49most softwares that release their MCP,
1:36:52they will have a URL exactly like this
1:36:53for their software.
1:36:54>> Uh and additional context to you guys,
1:36:56Beehiiv is obviously an email platform
1:36:58and it was one of the first, if not the
1:36:59first, with an MCP connector from from
1:37:01my knowledge. So.
1:37:03>> Yeah, I've been using MCP. I love it.
1:37:05So, I'm going to show you like we'll
1:37:06we'll grab this and this URL. So, we're
1:37:09back into Claude and we'll go paste and
1:37:11then add.
1:37:13And then now it's given Beehiiv to us as
1:37:15a connector. And then now let's go into
1:37:17like a a fresh session. I'm working out
1:37:20of my OS folder, which I'm going to take
1:37:22out the hood in there as well in a sec.
1:37:23And then I'm just going to say,
1:37:25"Could you please go into Notion and
1:37:28find my most recent newsletter?"
1:37:30And then
1:37:33I haven't posted it yet. So, we need to
1:37:34create some subject lines. I want you to
1:37:36then just go into Beehiiv with the
1:37:39Beehiiv MCP and find like all my subject
1:37:41lines and then pull in like all the data
1:37:44so you can see which ones performed the
1:37:45best and use that to help write like 10
1:37:49new options for the most recent
1:37:51newsletter.
1:37:52Send that off.
1:37:54And you can see like with my prompts,
1:37:56there's no real structure or formatting
1:37:57to them. It's just like
1:37:59plain natural language.
1:38:00>> Well, that's Whisper. It obviously
1:38:01breaks it down and kind of
1:38:02streamlines and optimizes the ones that
1:38:04are working.
1:38:04>> Yeah.
1:38:05But I know like some people are like
1:38:08worried that they don't know how to
1:38:08prompt it correctly, but it's like
1:38:11I don't know. I feel like prompting used
1:38:12to be this thing people prompting is a
1:38:14skill.
1:38:14>> becoming less and less. Yeah. And guys,
1:38:17anything that you can think of
1:38:19with this Beehiiv thing, I mean, it
1:38:20could do cohort analysis. We're going to
1:38:22do subject lines. I mean, it could look
1:38:24underneath the hood and see who's on
1:38:27your list. And if you want to say, you
1:38:29know, I have this cohort of operators or
1:38:32athletes and it can extract lists of
1:38:34those people. And do they even have this
1:38:37crazy, I don't know if you're using it,
1:38:38like the UTM tracking where you can also
1:38:41see like on an individual one-to-one
1:38:43basis, this person with this email
1:38:46has clicked on
1:38:48these specific emails. And like you see
1:38:50what percent of time they open up. You
1:38:51could see the source of where they came
1:38:53from from an attribution perspective. If
1:38:55you're doing like lead magnets, you can
1:38:57then see like, oh, they came from the
1:38:59George Heaton lead magnet. And then you
1:39:02could start evaluating like, okay, this
1:39:04lead magnet was the best lead magnet.
1:39:06And why was it the best lead like Dude,
1:39:08the rabbit hole is absolutely incredible
1:39:11on this.
1:39:12>> Yeah, 100%. I've been using the Beehiiv
1:39:13MCP as well for like so much stuff cuz I
1:39:16like have my newsletter that I write
1:39:18each week and it's like a big part of
1:39:19like what I do and then where I add
1:39:21value to to my audience. And yeah, it's
1:39:24it's just so good for like subject
1:39:26lines, like data analysis. I was getting
1:39:28my entire Beehiiv list into Resend to
1:39:31send some like
1:39:32like flows and stuff.
1:39:34And it just like one click moved it in
1:39:37for me. Like I just said to Claude like
1:39:40create the Resend audience from Beehiiv.
1:39:41And it just like did it for me with the
1:39:43MCP. Like just so handy. You can see
1:39:45here like first kind of in its first few
1:39:47loops around. We've said like go into
1:39:49notion, find my most recent newsletter
1:39:51first, and then cuz it needs subject
1:39:54lines to write for a newsletter, right?
1:39:56You can't have a subject line from you
1:39:57don't have an email. Yep. So it's going
1:39:59to do that first, and it's going through
1:40:00the loop right now to try and find that
1:40:01email. You can see what it's doing.
1:40:04Hey, while while that's cooking, I
1:40:06wanted to just like we can cut this in,
1:40:08but I wanted to show one other skill I
1:40:09built, but I'm going to do like a
1:40:1130-second version, so it'll just be
1:40:12super quick. I just want to blow
1:40:13people's minds here real quick. I don't
1:40:15know if you're ready for this one. Let's
1:40:16do it. So this is a product shoot skill.
1:40:18I actually built this for a buddy's
1:40:19company, and
1:40:21we have this skill so that the brands
1:40:24like Teva stones, we got like a Teva
1:40:25product shoots the skill I built. So I
1:40:27would just go literally I'd run Teva
1:40:29product shoot, and I would just get the
1:40:31reference image of the product. So like
1:40:34this cabinet for example, which can just
1:40:36be like iPhone images of the product.
1:40:39And then I would run the skill, point it
1:40:41at the image, and then it built it it
1:40:43would generate all of the
1:40:46production-ready images for the for the
1:40:48thing. So so this
1:40:49little cabinet you see here, that's the
1:40:51reference image. That's not the final
1:40:53one. Look at this.
1:40:55That's took that one that you just saw.
1:40:58And that's built it out into a full like
1:41:00editorial thing. In the same office, you
1:41:02can see here. Look at that. Ready? Now
1:41:03it's like a different angle with someone
1:41:05like touching it. And then
1:41:08it's got like a close-up of the veining
1:41:09on the marble.
1:41:12>> And basically the input output is just
1:41:13giving the the raw like basically CAD
1:41:15file, and then you have a very very
1:41:17detailed skill that outputs whatever,
1:41:19five or six of these things that could
1:41:20be like on the PDP page or something.
1:41:22>> Yeah. And this is another one. So this
1:41:23was the reference image we gave it, and
1:41:24then these are like the what it what it
1:41:26built for the like
1:41:28editorial. Like just insane.
1:41:32Look at these close-up ones. Like it is
1:41:34so insanely detailed.
1:41:36>> Yeah, and I I assume that this was
1:41:38filtered through the Hexfield MCP or
1:41:40not?
1:41:40>> Yeah, yeah, this was done with Hexfield
1:41:42MCP to generate the images. And it used
1:41:44I think like Nano Banana Pro as the
1:41:45image model, which is Google's one,
1:41:47through Hexfield though.
1:41:49So, that was like one of my favorite
1:41:51skills I've ever built.
1:41:53And then we'll cut that back to the
1:41:54Behive example. It pulled out and pulled
1:41:56out like some of the patterns that win.
1:41:57Like you can see it's got like don't
1:41:58stack two names, your two worst lines
1:42:00both jammed two tools into the name.
1:42:03Bold stands I plus bold stands wins. I
1:42:06plus boring task doesn't. I want my
1:42:07don't want my business to be trapped in
1:42:09Claude. Got a great one. I automated Mad
1:42:11Ads in 20 minutes even though I think
1:42:13that's a good subject line. Didn't do so
1:42:14well. But you can see it's pulled the
1:42:16gist of them. It's got like written out
1:42:18some for this new newsletter that we've
1:42:19got.
1:42:20And then I can now say as well
1:42:24and it see it's got that pattern as well
1:42:25cuz I usually put this little emoji at
1:42:26the start of all my subject lines. So,
1:42:28I'm just going to say to it to it now.
1:42:29I'm going to say like
1:42:31Great, can you please build a uh subject
1:42:33lines skill that always goes into Behive
1:42:37to check the best performing latest
1:42:40subject lines and historic as well and
1:42:42then uses that data to help write the
1:42:46subject line for the latest newsletter.
1:42:49You can see it's cooked away here. It's
1:42:50created the skill.md for us. And we can
1:42:54see that it's got the name subject lines
1:42:56description
1:42:57telling it like what to do, which is
1:42:59going to be on like the little spine of
1:43:00that book.
1:43:01>> And that that name the subject lines is
1:43:03the slash and then it's the the drop
1:43:05down menu and subject lines comes up.
1:43:07>> Yeah, yeah, exactly. So, I'll show you
1:43:08that in a second. You can see it's
1:43:09written this doc for itself. Hey, step
1:43:11one, load the tools.
1:43:13Step one, pull Behive performance data.
1:43:16Step three, find the draft and the hero
1:43:19story in Notion.
1:43:21Then write 10 ranked options.
1:43:23Output structure, use this shape when
1:43:25giving them to the reader. It's so
1:43:27simple.
1:43:28And then we can basically like
1:43:30that's saved the skill to our .Claude
1:43:34folder which I'll touch on in 2 secs,
1:43:36but let let's for this example, let's
1:43:39tell it
1:43:40Remember that demo folder that we built
1:43:42earlier where it had like our Claude.md
1:43:44and stuff in it? Let's get the skill in
1:43:46there so you guys can see how this looks
1:43:48when we like continue building out.
1:43:50>> And that's adding in like the brand
1:43:51voice and the stuff like that that we
1:43:52talked about earlier?
1:43:53>> Yeah, that project.
1:43:54>> Yep.
1:43:54>> Yeah, yeah. So I might say like, "Could
1:43:55you please find the demo folder on my
1:43:57desktop? And then I want this skill to
1:43:59be saved in that demo folder."
1:44:02Okay, so now it's saved it to our demo
1:44:05folder. So I I I want to just keep
1:44:07coming back to this example here cuz
1:44:08this is like our folder that we've been
1:44:09building on, right? We started with the
1:44:11Claude.md, we added the context in, the
1:44:12memory in.
1:44:14Now we can see it's created this little
1:44:16folder for us called the .Claude folder.
1:44:18>> Why's that dark?
1:44:19>> So
1:44:20the you see it's got the dot in front of
1:44:21it? Any folders that you have a dot in
1:44:24front of are hidden. So if we go command
1:44:26shift dot on the Mac, like period, it
1:44:29hides it.
1:44:30So by default when I'm working in my in
1:44:32my normal project, I don't see those
1:44:34folders unless I run that little command
1:44:36shift dot command. It just hides them. I
1:44:38just just for like workspace
1:44:39cleanliness.
1:44:40>> Yep.
1:44:40>> And then that's kind of like Claude's
1:44:42settings. And the same way Claude like
1:44:45when it's loading into a session in this
1:44:46demo folder, it will always know to load
1:44:49in Claude.md, it will also always check
1:44:52that .Claude folder as well. So quite
1:44:54often what I'll do in my projects is I
1:44:56will put all this inside the .Claude
1:44:58folder as well.
1:44:59So that way like I can just have that
1:45:00hidden and then I just have in my
1:45:02working folder
1:45:04that's kind of tucked away and I can
1:45:05just have my actual files that I care
1:45:07about. So if we open this up, it it
1:45:10created this little folder called skills
1:45:12and then it's added in our subject line
1:45:13skill
1:45:14with our skill.md.
1:45:16So that's in there nicely. We can open
1:45:18that up and there we go, we have our
1:45:19name, description, and contents with the
1:45:21little step-by-step guide for the
1:45:23subject lines. And then now all like
1:45:25these contexts and and skills and stuff,
1:45:27I just like to leave them in this doc
1:45:28cloud folder and then that's kind of
1:45:30like hidden away within the project.
1:45:32>> Yeah, and just uh just to tie a bow on
1:45:34that throughout you know, for everybody
1:45:35out there. So, it's basically you have
1:45:37the skill. It's the SOP on the actual
1:45:40structure on how you are now going to
1:45:43build subject lines moving forward
1:45:45tapping into the MCP with B Hive for the
1:45:47most optimized results and you can keep
1:45:49iterating and optimizing that actual
1:45:51skill.
1:45:52>> And you can change skills too. So, like
1:45:54my I've got a skill called YouTube
1:45:55publish workflow and what that does is
1:45:58if you open up the skill.md it literally
1:46:00tells it step-by-step. Hey, step one use
1:46:02the YouTube title skill. Hey, step two
1:46:05use the YouTube thumbnail skill. So, you
1:46:07can see you've got like an almost like
1:46:08an orchestrator skill that will like
1:46:10call the other skills in sequence. So,
1:46:12it's opening up book one for YouTube
1:46:15workflow and then it's telling him, all
1:46:16right, now you've got to get these four
1:46:18skills and he's like, okay, and then he
1:46:19just adds the other books and he'll keep
1:46:20working. So, like you can
1:46:22create some insane workflows.
1:46:24>> For people out there that worry that
1:46:25your brain is going to be irrelevant at
1:46:27some point, I mean, you just touched on
1:46:29something that is very very relevant is
1:46:31like
1:46:32the strategic mind is still very very
1:46:34important to understand the sequencing
1:46:36of how to look at things. Cuz some
1:46:38people want
1:46:39some people want to look at thumbnails
1:46:41before they do titles and without kind
1:46:42of giving it that input it would not
1:46:44know that. So,
1:46:45>> The chaining of skills that's very very
1:46:47interesting and that would just live in
1:46:49that core skill file and it just would
1:46:51say, check this, check this, check this
1:46:53in sequence.
1:46:53>> they're just separate ones. I'll show
1:46:54you.
1:46:56So, in my doc cloud folder with my OS
1:46:58folder, I literally just have these
1:47:00skills here. And for example, the I
1:47:02mean, that ads analyst one I ran earlier
1:47:04was a was a orchestrator skill. It
1:47:06called the ad creative analysis skill.
1:47:07It called the like creative strategy
1:47:10skill. It called a bunch of them. And
1:47:12then my YouTube publish workflow skill
1:47:16the skills that it's orchestrating are
1:47:18not within the folder. They're They're
1:47:20all stand-alone skills. So, I can just
1:47:22open this up and show you this skill.
1:47:25YouTube publish workflow literally says
1:47:27this is an orchestrator skill, use these
1:47:30other four skills in sequence.
1:47:32You can see how simple that is.
1:47:33>> Insane. I think we've been looking at a
1:47:35lot of folders and talking about a lot
1:47:37of things. Think that there's a lot for
1:47:38people to wrap their brain around. I'd
1:47:40love to kind of zoom out and kind of
1:47:41look at like your machine, your
1:47:43operating system. Talk about like folder
1:47:45structure and just see from a macro
1:47:47perspective like how everything is
1:47:49organized. So, then people can see what
1:47:51it looks like when you put in a [ __ ] ton
1:47:52of hours and kind of what the end looks
1:47:54like. So, I'd love to just give you the
1:47:55floor and maybe do like a live
1:47:56walk-through of what everything looks
1:47:57like on your set.
1:47:59>> 100%. So, the way that I run things, and
1:48:02this is like the the kind of way I've
1:48:03landed after months and months of use
1:48:05and trialing lots of different setups.
1:48:08>> Guys, real quick. Everything that Remi
1:48:09is setting up, the loop, the folder
1:48:11structure, the entire system, he made a
1:48:13plugin that does it for you, built
1:48:15around your life and your business. It
1:48:17interviews you, then connects your
1:48:19email, your calendar, your docs,
1:48:20everything. I am not technical. I did it
1:48:23myself. 90 minutes, one sitting, done.
1:48:25Sign up for the newsletter below and get
1:48:2750% off at launch. Enjoy the rest of the
1:48:30episode.
1:48:31>> I originally used to like kind of create
1:48:33these agents. Like I would have like
1:48:35really granular niche agents. Like I
1:48:37would have like a you know, I mean at
1:48:39one point I might have even had like a
1:48:40subject lines agent.
1:48:42>> Jesus.
1:48:43>> And I mean when I say agent, you know,
1:48:44folder. Like I would I definitely have
1:48:46like a you know, marketing like
1:48:48ads folder. And then I would kind of
1:48:50just like open up those folders to work
1:48:52in, if that makes sense. If I was doing
1:48:54ads, I would open up my ads folder and
1:48:55code code and like scope it to that
1:48:57folder and then work out of there. But
1:49:00the more I I
1:49:02build and work, it's like it's just such
1:49:03an annoying inefficient way to have
1:49:05things. I just run everything out of
1:49:08this OS folder. And
1:49:10for some reason as humans we love to
1:49:11like give these agents like treat them
1:49:14like humans, you know, give them human
1:49:15names and like give them like roles and
1:49:17things.
1:49:17>> Marty, we got a we got Marty there,
1:49:18right?
1:49:19>> Marty, yeah, exactly. Like people people
1:49:20for some reason they they they they want
1:49:22to have like a head of marketing agent,
1:49:23all this stuff. And it's also what gets
1:49:25a lot of clicks and views on Instagram.
1:49:27You know, people go like, "I've got my
1:49:28team of agents doing this." And it's
1:49:29like, really they're just
1:49:32folders. And the more folders you have,
1:49:35and more specialized, the more like
1:49:36different stuff there is to worry about.
1:49:38And the models are just so smart now
1:49:40that I just run everything out of the OS
1:49:42folder, and then
1:49:43I build really, really good skills to
1:49:45teach at the processes. And this is my
1:49:47OS folder here, right? So, I have
1:49:50all of my like
1:49:52This is all my stuff for my like, you
1:49:54know, education consulting.
1:49:56This is stuff for for Comfort work I'm
1:49:58doing there. And this is just another
1:49:59like thing just for for some notes and
1:50:01stuff.
1:50:02>> So, so let's take a step back. So, if
1:50:04somebody's looking at this, and they
1:50:05want to look at their folder structure.
1:50:07Let's just say for example, for me, like
1:50:09I would have Iconic, and I would have
1:50:11Open Residency. Like these are the big
1:50:12pillars in which I want to work in.
1:50:14>> 100% Yeah. So, you got it exactly right.
1:50:15So, you can see like my my folder
1:50:17structure is literally like I've got
1:50:18like my overarching OS folder. OS just
1:50:21stands for operating system.
1:50:23And then I have, you know, any like any
1:50:26businesses or pillars. So, yeah, for
1:50:28you, Iconic and Open Residency. You
1:50:29might even have another one for like
1:50:32You might have a folder just called
1:50:33personal for like your health stuff that
1:50:34you wanted to do.
1:50:35>> Yeah.
1:50:36>> And I've got another one that I like to
1:50:38have. I just have a folder called
1:50:39active, which is just like one-off
1:50:41projects and experiments that don't
1:50:42really fit into like one of the pillars.
1:50:44>> And if potentially becomes something
1:50:45substantial, maybe you'd move it into
1:50:47one of those other portfolios.
1:50:48>> So, for like I created a folder for this
1:50:50podcast to prep. And I just added it.
1:50:52It's just sitting in the active folder.
1:50:54It's loose. It's just like a separate
1:50:55project that I'm working on.
1:50:57And then within these like overarching
1:50:59pillars, then I'll have like the folders
1:51:01within there, like content, newsletter,
1:51:02website. So, for example, in my real OS
1:51:05folder,
1:51:06you see iOS by me. We go in, and I've
1:51:08got like content, finance, legal,
1:51:12newsletter,
1:51:13website, marketing, and I used to
1:51:17open Claude. So, I would in the past
1:51:20have Let's just say I wanted to work on
1:51:21marketing. What I would have done is I
1:51:24would have opened up the marketing
1:51:25folder
1:51:26and then started working in there for
1:51:28all my marketing stuff.
1:51:29>> And what's the right way to do it?
1:51:30>> I mean, there is no right or wrong way,
1:51:31but I I The reason I was doing that and
1:51:33my thought process was that then it's
1:51:34like we're scoping the agent to just
1:51:36marketing work. But they're so like the
1:51:38agents are so smart, the models are so
1:51:40good in their context windows, which is
1:51:41like the bar chart, like how much you
1:51:43can fit in it, is just getting bigger
1:51:45and bigger, that I just run everything
1:51:47out of my OS folder.
1:51:48>> That's under the notion that the OS is
1:51:50very macro, and it's going to pick up
1:51:52all the stuff underneath it as opposed
1:51:54to going
1:51:54>> Yeah.
1:51:55>> granular.
1:51:55>> And what I would do, right? So, let's
1:51:56just say like for my new like a
1:51:58newsletter or or marketing Let's say
1:51:59we're creating a new batch of ads. I
1:52:00would sort of have OS folder open
1:52:02working, and then I like to just go like
1:52:04this. I'll tag the folder. You can tag
1:52:06folders.
1:52:07And then I would go, you know, AI
1:52:11with
1:52:12Remy
1:52:14/
1:52:16marketing.
1:52:17So, it's got the marketing folder.
1:52:18>> there has to be some muscle memory in
1:52:19there and
1:52:20and knowing what your kind of, you know,
1:52:22folder structure is. Yeah. I would love
1:52:23for you to walk through right there,
1:52:24just the ads, the slashes. Like, what
1:52:26are all those kind of like quick
1:52:27shortcuts that people could do to tap in
1:52:28there?
1:52:30>> I mean, it's going to be slightly
1:52:31different depending on what harness
1:52:32you're in, but most of them like Codex,
1:52:34Claude Code, pretty much all of them use
1:52:36a slash command and an at symbol type
1:52:39system. So, slash command just brings up
1:52:41all your skills and like commands
1:52:42available to run.
1:52:44As you can see, we've got a lot.
1:52:45And but it also just brings up like
1:52:47settings that are baked into Claude.
1:52:49Like, for example, if I want to change
1:52:50the model, the LLM, the brain behind
1:52:52this loop, I can click the slash model,
1:52:55and then it's going to bring it up
1:52:56there.
1:52:57>> And for context, too, for people, I just
1:52:58want to double click on everything to
1:52:59give people 100% clarity on stuff, too.
1:53:01If you're going to slash skill,
1:53:03is that like tag teaming it in? What
1:53:05happens if you don't tag team in it tag
1:53:07team it in. Is it going to automatically
1:53:09tap into the skill with its own brain or
1:53:11do you have to be very intentional in
1:53:13making sure that you're tapping in the
1:53:14skill?
1:53:15>> Sometimes I'm intentional. If I say
1:53:17write me a a hook for my Instagram video
1:53:21and I've got a hook skill, it's just
1:53:22it's going to pick it up.
1:53:23>> Yeah.
1:53:23>> But like for example the ads analyst
1:53:25one, if I just dumped in the ads library
1:53:26link and ran it, it would be like what
1:53:28the [ __ ] do I do with this?
1:53:29>> So naming conventions are important but
1:53:31obviously like really having muscle
1:53:34memory in what your folder structure is
1:53:35so you know when to tap into certain
1:53:37stuff. It's basically just like you have
1:53:3950 employees, you should know the names
1:53:40of your employees. Kind of
1:53:41>> Yeah, yeah, yeah, yeah, yeah, yeah,
1:53:42yeah, exactly same thing. I mean like
1:53:43I've got a pretty big setup and it's not
1:53:45that hard to like work it out. Like I
1:53:46mean even so like slash commands is for
1:53:48your skills and like other little
1:53:50setting stuff, mainly skills. But the at
1:53:53I find is like when you want to tag
1:53:54files or or folders. So
1:53:57like all documents, you know, for
1:53:59example. Um
1:54:00>> And should that be under the notion that
1:54:01if you tag in a folder, it is going to
1:54:03pick up all of the files within that
1:54:06folder?
1:54:06>> Yeah. Yeah. So, you know, I I wouldn't
1:54:08even really have to remember. I think if
1:54:10I just go at marketing, right?
1:54:12I want to do marketing. It brings up
1:54:14straight away my marketing folder. Hey,
1:54:16it's Remy marketing.
1:54:18I can select that. And now if I give it
1:54:20my task for marketing like write me 10
1:54:23new ads for my upcoming workshop.
1:54:26Then you will see what it will do is it
1:54:29will then
1:54:30start to like use that folder to work in
1:54:33but without us having to actually open
1:54:35that folder up to work in. Does that
1:54:37make sense?
1:54:37>> Yeah, be your your basically uploading
1:54:40the OS folder which is macro and it's
1:54:42getting everything under it but then
1:54:44you're tag teaming in
1:54:45>> Yeah.
1:54:45>> the marketing folder to take like an
1:54:47extra good look at the marketing folder.
1:54:48>> Yep, that's exactly it.
1:54:49>> I don't know if I explained that good
1:54:50for you guys out there but I think it's
1:54:51a I'm I'm trying to make this at a a
1:54:54fifth grade level so people could really
1:54:55understand it. It's just like hey, go
1:54:57check this folder. It's probably where
1:54:58it is.
1:54:59>> I mean that's like that's that's what I
1:55:00meant about demystifying this all. It's
1:55:02like you hear like people make it
1:55:03Instagram videos and it's like my mom my
1:55:06marketing agent just did XYZ while I was
1:55:08asleep overnight. Just worked all night.
1:55:10And it's like
1:55:12If you don't understand this, you're
1:55:13like, "Oh, like wow, this is crazy." But
1:55:15in reality, they've I I I guarantee you
1:55:18they have a folder called like marketing
1:55:20and they have some skills in there that
1:55:22are marketing related and they're
1:55:23running those skills on them some kind
1:55:25of schedule. So, it's like every night
1:55:27do a report on Facebook ads and email to
1:55:28me so I wake up to a report. So so
1:55:31simple. It's just folder, skills,
1:55:33schedule tasks. And then But people like
1:55:35take that and I I don't know whether
1:55:37it's an ego thing cuz they they like to
1:55:39sound smart and sound cool or whether
1:55:41they're pushing some course and that
1:55:42they want you to think they're like some
1:55:43AI god. But in reality, it's like just
1:55:46as simple as the skill and a folder and
1:55:48a schedule task. But
1:55:50people will say, "I've got like
1:55:5250 agents working for me overnight."
1:55:54>> And that could be anything. It could
1:55:55literally be clean the inbox. You can
1:55:57say draft draft email responses.
1:56:00>> Send cold emails or like send LinkedIn
1:56:01connections, you know.
1:56:02>> Suggest emails that you want to
1:56:04unsubscribe to if you haven't opened up
1:56:05the last three four times.
1:56:07>> Yeah. So, it's like, you know, now
1:56:09If If I set this up and click create,
1:56:11now I've got an agent that is managing
1:56:14my email inbox for me.
1:56:16>> Insane.
1:56:17>> It's It's It's a lot simpler than you
1:56:19think. The last sort of concept that
1:56:20goes really hand in hand with this sort
1:56:22of folder structure I'm outlining is you
1:56:24have
1:56:25context, tools, and skills. You can have
1:56:27them at a global versus a project level.
1:56:30Easiest way to think of global is like a
1:56:31holding company. And then
1:56:34the project level is like the sub
1:56:35companies, which is kind of like how I
1:56:37have my setup. And
1:56:39you know, things at a at a OS level like
1:56:42a a global or a holding company level
1:56:44are things that apply to like every
1:56:46single company. So, if you're looking at
1:56:47this holding company example, you might
1:56:49have like context, tools, and skills
1:56:52that apply to like all of your
1:56:53companies. And then you might have some
1:56:56things that only apply to really really
1:56:58specific companies. So, I have
1:57:01like skills that live at a global level,
1:57:03which I can access anywhere. And I also
1:57:05have like skills that exist just within
1:57:07a folder. Like remember that one I
1:57:09showed you before with the product
1:57:10shoot, taper product shoot. That skill
1:57:12I've got kept in the folder for the
1:57:14taper stuff. Cuz like I'm never going to
1:57:16use that skill outside of that specific
1:57:19folder.
1:57:20>> It's so crazy cuz this is really just
1:57:21like this whole thing it's a personal
1:57:24OS, but it's really just a holding
1:57:25company.
1:57:26>> Yeah, the whole entire infrastructure of
1:57:27the folders is like a holding company.
1:57:28>> Yeah, yeah. Exactly. That's that's
1:57:29exactly the way to think about it. So,
1:57:32another example of like global versus
1:57:34project or holding company versus
1:57:36company specific is our cloud.md.
1:57:38Remember that?
1:57:38>> Yeah.
1:57:39>> So, I have one that I have in my OS
1:57:42folder, my global cloud.md that I want
1:57:44to apply to like every single session
1:57:46I'm in.
1:57:47But I have a cloud.md in some of these
1:57:49sub projects. Like I'll give you an
1:57:51example right here. I'll open up the
1:57:52folders. So, I'm in my OS folder right
1:57:55now, like my real one. And you can see
1:57:57that my like
1:57:59you know, in my I guess you can call it
1:58:00like your settings folder,
1:58:02I've got my my cloud.md, my global one
1:58:04for my OS folder. And this one here
1:58:06contains like overall preferences about
1:58:08myself, you know, explains like who I am
1:58:11as a non-technical founder learning.
1:58:14That's not going to change no matter
1:58:15what project I work in. I'm always going
1:58:17to be the non-technical founder who's
1:58:18learning. So, I want that applied
1:58:20everywhere. I've got in here there's a
1:58:22loads of lessons that I've built up over
1:58:25months of years.
1:58:27Uh I've got
1:58:27>> I saw in there like core principles,
1:58:29core values. That's a great thing.
1:58:30>> is one I have as well, boil the ocean.
1:58:32The marginal cost of completeness is
1:58:33near zero. Basically just tells like the
1:58:35age of the standard isn't good enough.
1:58:36It's holy [ __ ] that's done. Like just
1:58:38little things like that, right? And
1:58:41I want that applying to every session.
1:58:42But for example, when I was prepping for
1:58:44the podcast, I created this little
1:58:47podcast folder called
1:58:49>> podcasts.
1:58:50>> podcasts, right? And then in here, this
1:58:53is where I was just like work shopping
1:58:54and planning everything, right? You can
1:58:55see a bunch of files. This is just like
1:58:57a big like work work space essentially.
1:59:00But you can see within the podcast
1:59:02folder, I've got a Claude that MD. And
1:59:04if we open that up, that just contains
1:59:06like the North Star context for the
1:59:09podcasts folder.
1:59:10>> Yeah. So it's it's it's going to be
1:59:12filtered through the macro Claude folder
1:59:14for like your principles and how you
1:59:16think and then this is more like rubric
1:59:19and outline, yeah.
1:59:21>> So it's simple, ready? This workspace is
1:59:22Remi's base for podcast appearances.
1:59:25What this folder's for prep, demos,
1:59:27diagrams, post episode analysis.
1:59:30That's literally all it is. And then
1:59:33I just have some other folders in here I
1:59:35was working on stuff. But now when I'm
1:59:37opening up Claude to work, I am still
1:59:40working out of the OS folder and I would
1:59:41tag podcasts. Let's say I wanted to prep
1:59:44for this pod, I would tag podcasts. And
1:59:46then I would give it my task and I'd be
1:59:48like
1:59:49what's the outline for the open
1:59:50residency podcast?
1:59:53Even if I didn't tag podcast the folder,
1:59:55it would still go and find it, but I
1:59:57just like to do it for nice like
1:59:59cleanliness and peace of mind.
2:00:01Just to make sure.
2:00:02And you can see here it's opening it up
2:00:04and then it's working.
2:00:05>> down for everybody out there listening,
2:00:06you brought in the OS folder so it's
2:00:08getting the macro Claude and all the
2:00:10stuff there and now you're tapping in
2:00:12this skill of podcast and it's picking
2:00:15up the Claude MD file within there and
2:00:17anything and everything yep around
2:00:18there.
2:00:18>> And and like now loaded in the Claude
2:00:20that MD from the podcast folder. So
2:00:22you're kind of like stacking them on top
2:00:24of each other. Yep.
2:00:25And you can see this whipped up our our
2:00:27episode plan here, which is which is
2:00:29super cool. That's just giving an
2:00:30example of like global versus project
2:00:33level or holding company versus company
2:00:35level Claude that MDs, skills and tools.
2:00:40>> Amazing amazing context there. I'd love
2:00:42to just chew it for my own selfish
2:00:43reasons and I think for a lot of people
2:00:45out there they're probably are
2:00:46interested in this as well. Just like
2:00:47something that I've been working on and
2:00:49just how would you kind of construct it
2:00:51on a global verse project level is like
2:00:53me I'm like on a health journey. Let's
2:00:55just say I have my DEXA scan, my blood
2:00:57work, maybe I'm tracking my macros. This
2:01:00whole entire flywheel of all these kind
2:01:01of touch points, maybe I have my aura
2:01:03sleep score which has an MCP and can
2:01:05bring it in. Maybe eight sleep and bring
2:01:06it in.
2:01:07And and can bring it in. Like how would
2:01:09you architect that? Global verse project
2:01:11and then like just give me your your
2:01:12thoughts on that.
2:01:13>> Yeah, so I'd have my I guess holding
2:01:15company like
2:01:16OS folder. Yep. And then you know, you
2:01:18might have you'd have a folder within
2:01:19that for like open residency, another
2:01:21one for iconic. And then I'd probably
2:01:22have another folder of cuz it's a big
2:01:23pillar in your life like personal. And
2:01:25then within personal, I'd probably have
2:01:27a folder called like health.
2:01:29And then I would just in that health
2:01:31folder
2:01:32I would then just dump in all that
2:01:33stuff. Everything. Just dump it in like
2:01:35DEXA scans, whatever. Just just dump it
2:01:37in there.
2:01:38And then I would just create a Claude.md
2:01:40just for the health folder.
2:01:42And I would basically just say something
2:01:43along the lines of
2:01:44this is my health workspace where I'm
2:01:48going to be, you know, analyzing blood
2:01:50work and doing this and doing that. You
2:01:52might even then give it like your your
2:01:53health goals and put it in the
2:01:55Claude.md. Just any context that your
2:01:58agent might need when it's like when you
2:02:00want to work on health stuff.
2:02:01>> Yeah. And then that's literally it.
2:02:03>> Yeah, I think like again, to kind of
2:02:04take it a step further, something that
2:02:05I'm working on now is just like okay,
2:02:07you have your separate vendors. Maybe it
2:02:09takes X amount of time to book an
2:02:10appointment to go get your blood work,
2:02:12go get your DEXA scan, you know, maybe
2:02:14this is your preferred vendor to get the
2:02:15DEXA scan. You have your Gmail linked
2:02:18up. It knows every seven weeks to email
2:02:22the DEXA scan place to set up a DEXA
2:02:24scan so you stay on target and keep
2:02:26updating it. So there's a lot of ways.
2:02:28That's something that I'm working on.
2:02:29I'm super super excited about to get the
2:02:31reminders and automate the the different
2:02:33appointments.
2:02:34>> Yeah, I know a few guys that have like
2:02:36done crazy stuff with like health and
2:02:38blood work with Claude code and and
2:02:40agents.
2:02:41>> Yeah. For me it's simple cuz it's just
2:02:42like I saw my blood work that like I
2:02:44have to work on saturated fats and
2:02:45sugars and like I just now work through
2:02:48the lens of what food do I eat and is it
2:02:51bad and obviously calories and protein
2:02:53etc. but like what does it look like on
2:02:55those kind of two realms?
2:02:56>> Yeah.
2:02:57>> Amazing. Definitely be talking about
2:02:59that kind of health whole entire
2:03:02flywheel that I'm setting up. I think a
2:03:03lot of people would be interested in
2:03:04that. What is like the honest kind of
2:03:06like monthly cost of doing all of this?
2:03:09Obviously you talked about the $100
2:03:10maybe $200 a month if you want to get
2:03:12the bigger plan on your core LLM. How
2:03:15much are you spending a month on kind of
2:03:16everything?
2:03:17>> Oh, everything all up? About a
2:03:22I mean about
2:03:24a thousand a month.
2:03:26But that's across like loads of tools
2:03:29and you know I'm I it's my job to be on
2:03:30top of this. I'm I'm testing stuff. I
2:03:32got subscriptions I forgot about from
2:03:33months ago. Like I'm Yeah, I I know my
2:03:36bill's about a thousand bucks for AI
2:03:37stuff.
2:03:37>> You need to set up a schedule to cancel
2:03:39the subscriptions that you're not
2:03:40>> You should bro, come on. Come on, man.
2:03:41Come on, dude.
2:03:42>> I'm on a schedule. I should have I
2:03:43should have one once a month that goes
2:03:45to my email receipts and and notifies
2:03:46me. I should.
2:03:47>> The email unsubscribe thing is is a
2:03:49game-changer.
2:03:50>> Yeah. So for people out there listening,
2:03:52so $12,000 for basically a year for I
2:03:55don't know 100 employees.
2:03:56>> I mean I think it's it's it's a little
2:03:58bit interesting because all of these big
2:03:59AI companies right now like OpenAI,
2:04:02Anthropic, they're all like none of them
2:04:03are profitable. I mean maybe Anthropic
2:04:05just got profitable. I have no idea. Yet
2:04:07to confirm. None of them are profitable.
2:04:08I don't know if you knew that. They are
2:04:10all running at massive losses because
2:04:14you know, the they on these plans you're
2:04:16buying like the free plan, the pro plan,
2:04:19the the $300 a month plan, they're
2:04:22heavily subsidizing the real compute
2:04:24cost of of
2:04:26like the AI usage, right? So I think on
2:04:30a $300 a month Claude plan, you are
2:04:33actually getting about $6,000 worth of
2:04:35token usage. But they are subsidizing it
2:04:38so heavily because they're all trying to
2:04:40get users. They're just trying to
2:04:41capture mass market share. And Anthropic
2:04:43are like really now making a push to IPO
2:04:45and become profitable. They're starting
2:04:47to like They've kept their prices the
2:04:48same, but your usage is going down and
2:04:50down and down getting squeezed. They've
2:04:51just released their latest model. I know
2:04:53like maybe if you're watching this, it's
2:04:54something new, but they've just released
2:04:56Fable 5. And they've essentially said
2:04:59like this is our new most powerful
2:05:00model.
2:05:00>> It's
2:05:01>> not not included in your plan. You have
2:05:03to pay the real cost of it. Which is
2:05:05like
2:05:06disgustingly high.
2:05:08>> That's a great That's a great new rev
2:05:10stream for them to print. I'm sure
2:05:11they're charging some sort of markup on
2:05:13it, but that's like
2:05:13>> Yeah.
2:05:14>> for anybody that really wants the best
2:05:16model, they're going to pay for it. I
2:05:18feel like people are going to run that
2:05:19up.
2:05:19>> Yeah. So it'll be interesting to see
2:05:20what happens over the next sort of year
2:05:22or two as these companies like they
2:05:24can't keep running at a loss forever.
2:05:25They're going to have to start to bring
2:05:26the costs up.
2:05:28So it'll be interesting to see what
2:05:29happens. I think there's going to be a
2:05:30massive market for like open source
2:05:32models, basically models you can run
2:05:34yourself in your by like you know,
2:05:36couple Mac minis and and run your own AI
2:05:38models on there.
2:05:40I think there'll there'll always be
2:05:41options for like lower cost AI. There
2:05:42always will be, but it's just been
2:05:44interesting to see these big companies
2:05:46like what happens with the price of like
2:05:48your your plans.
2:05:50>> I completely agree. We've only talked
2:05:52you know, I mentioned the Kia. I I think
2:05:53that there's only Lamborghinis and Range
2:05:55Rovers right now. Eventually there's
2:05:56going to be some Kias and Hondas and all
2:05:58that is going to
2:05:58>> There's a few. There's
2:05:59>> There there's going to be a lot more
2:06:00options I feel like on the lower end. I
2:06:02don't know how they're going to figure
2:06:03it out, but definitely will be.
2:06:04>> Before we move on to the next thing from
2:06:06a a folder structure perspective and you
2:06:07know, global versus project, anything
2:06:09else big that you want to touch on there
2:06:11or do you want to move on?
2:06:13>> That is pretty much like all you need to
2:06:16know. And you know, I wouldn't get like
2:06:18keyed up about having like the perfect
2:06:19folder structure. It grows with you over
2:06:21time. Just start out with with Create a
2:06:24folder called OS. That's your That's
2:06:25your first step. And then just you know,
2:06:27create your code that I'm doing the
2:06:28context in.
2:06:29And
2:06:31those subfolders and like you know,
2:06:33holding com- like companies within
2:06:34there, that will just come over time as
2:06:36you start to like do real work. Like, I
2:06:38wouldn't go out and like try and get the
2:06:40perfect structure.
2:06:41I wouldn't let that hold you up from
2:06:42just starting and and getting work done
2:06:44with these agents.
2:06:46>> Just start.
2:06:47The two layers of AI in business. I'd
2:06:49love to kind of walk through your thesis
2:06:51on the different layers and how you look
2:06:53at it, specifically from a business
2:06:54perspective.
2:06:55>> So, the way I see it right now, you
2:06:57almost have like two layers of of AI
2:07:00within within the business, right? It's
2:07:01like you have your personal operating
2:07:03system, which is pretty much everything
2:07:05that we've covered today.
2:07:06This in my mind is like an extension of
2:07:08you. It's like you're going to become
2:07:09like an AI-augmented founder. Like, the
2:07:11the the way I think of it is it's like
2:07:13imagine a dude with like a robot arm.
2:07:14It's like an extension of of him and now
2:07:16he can like climb walls with his arm and
2:07:19do all this crazy stuff. And that's like
2:07:21this is your personal operating system,
2:07:24personal AI OS is kind of where you live
2:07:25in every day. This is like this Claude
2:07:28code thing I'm setting up with these
2:07:29folders where I now complete my entire
2:07:31work day from within there.
2:07:33And
2:07:34you know, that kind of handles all my
2:07:36tasks and my grunt work.
2:07:38But
2:07:39if you think big picture,
2:07:41let's just say you wanted to have an
2:07:42agent that managed your customer service
2:07:44or customer support.
2:07:46The it would be the wrong place to have
2:07:49that on your computer as the founder or,
2:07:51you know, your customer support team.
2:07:53Like, it probably wouldn't be
2:07:55make the most sense for them to have
2:07:57that on their computers.
2:07:59You kind of like that that's where this
2:08:00second layer comes in, which I called it
2:08:02more like your AI business
2:08:03infrastructure. And that's like
2:08:06that that's where some of those other
2:08:07agent frameworks like Hermes and Open
2:08:09Claw really come into play.
2:08:11>> More advanced frameworks.
2:08:12>> More advanced frameworks. And they're
2:08:13the kind of things like that's why
2:08:14everyone's running that's why Mac minis
2:08:15are flying off the shelves because
2:08:17people just wanted like a computer that
2:08:18was
2:08:20not theirs. Almost like a company Mac
2:08:21mini to set some of these agents up to
2:08:23run on.
2:08:24And but if you if you look under the
2:08:26hood at these agents, so right now like
2:08:28I just remote accessed into my Mac Mini
2:08:31that I have at home. And you can see
2:08:33here I've got like a a Hermez agent set
2:08:35up. And this is like the file system
2:08:38you're looking at right now is not on my
2:08:39laptop. This is the files on my Mac Mini
2:08:41at home. And you can see
2:08:43>> This is like the cloud.
2:08:44>> Yeah. Yeah. So, just like I have the OS
2:08:46folder that I work out of, the the the
2:08:49Hermez agent is literally just a folder
2:08:50as well. See dot Hermez? Yep. We open
2:08:52that up.
2:08:54But, it's just like a bunch of folders
2:08:55and and markdown files. You know, like a
2:08:56soul.md which tells it how to behave.
2:08:58Like
2:08:59It's all the same principles
2:09:01that we have just gone through today.
2:09:03So, when you eventually decide you want
2:09:04to like graduate to start using like
2:09:06something like an open core or a Hermez
2:09:07agent for different stuff,
2:09:09to if you want to configure these agents
2:09:11properly, still the same thing. Context,
2:09:13tools, skills. Nothing changes. It's
2:09:15just like they're just slightly more
2:09:17advanced frameworks that you that work
2:09:19better to run like
2:09:21more autonomously on another machine, if
2:09:23that makes sense.
2:09:23>> Yeah. I feel like there's a a huge
2:09:25opportunity for I think
2:09:28first steps is to become acquainted and
2:09:30understand the tools and do your own
2:09:32kind of personal OS. But, it begs to ask
2:09:34the you know, the question that
2:09:36everybody's asking next is like
2:09:38okay, who owns those files? That's for
2:09:40the individual person. Is it for the
2:09:42person? Is it for the role in the
2:09:44company? And then is there like a
2:09:46separate infrastructure for a company?
2:09:49Like how is there, you know, someone
2:09:51works for a company, they're setting up
2:09:52all of these skills,
2:09:54>> Yeah.
2:09:54>> then what happens if they leave the
2:09:55company? Where do the skills go?
2:09:56>> Yeah, it is like cuz this is so new,
2:09:59it's so messy right now. Like it is it's
2:10:01really this is all really simple at a
2:10:03personal level, right? Like you build
2:10:05skills, you connect your tools, you have
2:10:06your folders, and they just live on your
2:10:07computer where you work. It's simple.
2:10:09But, yeah, it it it does start to get
2:10:11messy when you for example like trying
2:10:13to work with a team and you know,
2:10:15someone builds like a stock forecasting
2:10:17skill, and then now,
2:10:19you know, another a member wants to use
2:10:21it, but it's on their computer, and it's
2:10:23like, "Well, how do you give it your
2:10:25team member access?" Oh, well, you could
2:10:27airdrop them the file, but then you've
2:10:28got two separate copies, and they're
2:10:29going to get out of sync fast if people
2:10:31are like making updates to them. So, to
2:10:34answer your question, like there's
2:10:35there's
2:10:36there's a few ways you can kind of
2:10:37bootleg it with like tools like GitHub
2:10:39and and Google Drive and stuff, but
2:10:41there's no like real clean simple way
2:10:44that I've found yet
2:10:46to to effectively like manage skills
2:10:49across like team members and companies.
2:10:52>> There's a massive opportunity for
2:10:53somebody out there to create the
2:10:55technology that has that connective
2:10:56tissue where kind of auto updates and I
2:10:58can share, for sure.
2:10:59>> There are some startups I've seen that
2:11:00are now starting to solve this with like
2:11:02shared skill management, and there's a
2:11:04There are a few coming out now. And I
2:11:06need to just sort of get my hands a bit
2:11:07dirty in that area of things. But, uh
2:11:10yeah, in terms of like skill ownership,
2:11:11like skills are like assets for your
2:11:12company. So, it's it's it's a whole
2:11:14'nother conversation as to whether like
2:11:17I think every employee needs to be the
2:11:18ones building them because, you know,
2:11:20they do the task the best. Like, they
2:11:22need to build the skills, but then it's
2:11:24like, if they're assets for the company,
2:11:26are you going to just let them walk away
2:11:28with the skills when they leave, or do
2:11:30you as a company own them? It's a
2:11:31difficult one. I can't say.
2:11:33>> Yeah, I feel like at a macro level, like
2:11:34the, you know, the Cloud MD file, like
2:11:36that should be macro and uniform, the
2:11:38stuff that doesn't change that we talked
2:11:39about. Like, that should be kind of
2:11:41across the whole entire company as like
2:11:43the core input in whatever folder
2:11:45they're working in for the company. But,
2:11:47then once you start kind of going
2:11:48downstream on the skills, I mean, that's
2:11:50where like creativity and taste and like
2:11:52how you prefer things to be done. Like,
2:11:55some people, you know, want to work in X
2:11:57program, other people want to work in in
2:11:58in this program. Some people, you know,
2:12:00need a link copy, some people need it,
2:12:01you know, are visual persons. So, I feel
2:12:03like that's
2:12:04it is very tailored toward the
2:12:05individual human.
2:12:06>> Yeah, I agree. But, you know, I I
2:12:08wouldn't get too
2:12:09caught up about it if you're just
2:12:10starting, you know? Just start Just like
2:12:12build your own skills first, and then
2:12:14worry about all the team management and
2:12:16other stuff like later. It's just cross
2:12:17that bridge when you're at it.
2:12:19>> Yeah, if you're out there listening, if
2:12:20you do even
2:12:2210% of what we talked about today,
2:12:24you're in the 0.000001%
2:12:27of people that have graduated from chat
2:12:29everything else past that. So,
2:12:32yeah, I think that's for another day.
2:12:34That kind of that layer two with the
2:12:36open claw and and the Hermes and such.
2:12:39>> Yeah. I mean, one of the thing I wanted
2:12:40to touch on too is like
2:12:42every company right now needs to like
2:12:43push to become AI native. Like every
2:12:45company needs to like put AI at the
2:12:47forefront of like all their operations
2:12:49or they're going to lose to a company
2:12:51that does. It's like it's simple as
2:12:52that.
2:12:53And
2:12:54like I've kind kind of been trying to
2:12:55solve some of these problems for
2:12:56companies I work with about the best way
2:12:58to go about doing this. And
2:13:00there's two ways to go about it and I
2:13:02and I and I think that one way is more
2:13:04correct than the other. So, the first
2:13:05way is like you go top-down, right? Like
2:13:08you bring in some AI specialists into
2:13:09the company to like start building
2:13:11stuff.
2:13:12And you know, that like they can go and
2:13:14interview each department and sit with
2:13:16the employees and like try and build
2:13:17skills on their behalf or try and like
2:13:19build agents to take roles. But it's
2:13:21like
2:13:23the employees the people that do those
2:13:25tasks, they know it so well in and out
2:13:28and there's so much nuance to every
2:13:29task. And if you want to bring AI
2:13:30specialists in to start automating this
2:13:32stuff, they would need to understand so
2:13:35deeply every task that everyone in the
2:13:37entire company does. They can never do
2:13:39that.
2:13:40So, you either need a lot of people. You
2:13:41need almost like one per one AI
2:13:44specialist for each employee if you're
2:13:46trying to go the top-down route where
2:13:47you build it and hand it to them and
2:13:48say, "Here, use it."
2:13:50I think the best way to get your company
2:13:52AI native is to
2:13:55get your employees trained up to be AI
2:13:57literate, get them building their own
2:13:58personal operating system. So, then you
2:14:00can the goal is not like fire entire
2:14:02team. You want to keep your best people,
2:14:04train them so they can become like 100x
2:14:06employees, and then you can go on hiring
2:14:08freeze and continue to grow without
2:14:09scaling a count. And
2:14:12the way you can start to automate is you
2:14:14you instead of from the top down, you
2:14:15automate it from the bottom up. So, you
2:14:17can train your employees to become AI
2:14:18literate, you get them to automate their
2:14:20own job, essentially. And then you
2:14:22eventually like have this like you know,
2:14:23if you think of like a a graph that's
2:14:25getting like saturated, you know, the AI
2:14:28like is saturated.
2:14:29>> Everyone in the company company everyone
2:14:30>> Yeah, exactly. That's probably like the
2:14:31best way to go about like trying to get
2:14:34your company AI native. And there's all
2:14:34sorts of things there that I think work
2:14:36well like
2:14:37putting incentives around it. Like you
2:14:39can do like build weeks and
2:14:40competitions, you know, like running
2:14:42Good idea. things like you know, best
2:14:43like skill of the month. Maybe you get
2:14:45your employees to start building skills.
2:14:47And then it's like the best one you win
2:14:48like a
2:14:49cash prize, $1,000. Like I think
2:14:51companies need to like just make AI like
2:14:53part of the culture. They need to have a
2:14:54an AI call each week where people are
2:14:56like sharing about what they've built
2:14:57and like really start to like bring that
2:14:59in as part of the company culture.
2:15:01>> I think it's a double-edged sword cuz
2:15:02some people, you know, the more they
2:15:04adopt, they feel like that they're
2:15:05digging their own grave to be fired. And
2:15:07that lends, you know, you have to talk
2:15:09about like what are actually skills that
2:15:11are going to do well in the new economy.
2:15:14I would disagree though on going two
2:15:16only. I would probably go a mix of path
2:15:18one and two. Like for me with Iconic
2:15:21like
2:15:22I had like a a six-page document that it
2:15:24was, let's just say, 60% of it was
2:15:27uniform in nature of like asking all of
2:15:29the probing questions. Like what do you
2:15:30you know, what tools do you use? What
2:15:32are the processes that you already have?
2:15:33What are your bottlenecks?
2:15:34>> Yeah.
2:15:34>> And then you're right, that other 40%
2:15:36was 100% tailored to that specific
2:15:40employee.
2:15:40>> Yeah.
2:15:41>> But with that being said, it's very
2:15:42tough. Like someone that's coming from
2:15:43the outside, they have to be very very
2:15:46versed in business to be able to pick up
2:15:49context fast. Yeah. Or else it's going
2:15:51to take a very very long time. But I
2:15:53think
2:15:53that's the first thing that every
2:15:55operator out there should do should be
2:15:56like the download of just like
2:15:59what AI tools do you use? And right away
2:16:02you're going to get a pulse check on
2:16:03like where everybody is at. Yeah. And
2:16:05like just seeing that for my own entire
2:16:07team, like that was a big shot in the
2:16:09arm to be like, "Wow, we got to like
2:16:11[ __ ] hurry up on this shit." Cuz with
2:16:13one, you will go faster. If you do just
2:16:15two, it's going to be slow. I think it's
2:16:16a mix between one and two.
2:16:18>> Yeah, I would agree with that. I would
2:16:19agree with that 100%. All right, this
2:16:20has been an amazing episode. Computers
2:16:21are down. Let's get into quick strike
2:16:23random questions we're just going to rip
2:16:25off.
2:16:25>> Let's go.
2:16:26>> Give me what you're thinking.
2:16:27>> So,
2:16:28first skill everyone should build.
2:16:31Email inbox triage skill.
2:16:34Run that bad boy on a schedule three
2:16:35times a day.
2:16:36>> Clean your brain.
2:16:37>> sorted.
2:16:39Too much noise.
2:16:40>> Biggest beginner mistake you see in AI.
2:16:43>> People
2:16:44being too perfectionist and too scared
2:16:46to start cuz they like don't want to
2:16:48screw something up. Like one of my
2:16:49buddies, Will, shout out if you're
2:16:50listening. They're like, "We don't want
2:16:52to use Claude, but like
2:16:53we're scared to even like type a prompt
2:16:55and we don't want to screw it up." And
2:16:56it's like, you know, cuz people kind of
2:16:58think it see it as a black box. They
2:16:59don't know what's going on under the
2:17:00hood. But it's like just the biggest
2:17:03beginner mistake is just being too kind
2:17:05of scared to start cuz you don't want to
2:17:06like screw things up and trying to be
2:17:08too perfectionist about the whole thing.
2:17:09>> Ladies and gentlemen, if you have made
2:17:10it this far and you don't take action,
2:17:13you are a loser.
2:17:14One thing you automated this week most
2:17:16people would never think to.
2:17:18>> My laundry.
2:17:20>> That's that's kind of crazy. Most
2:17:21overrated AI tool right now.
2:17:25>> I'd say any AI app builders. Lovable,
2:17:27Bolt, any of those. And you just
2:17:30just do it in Claude. Like just
2:17:32create a new I was just saying like
2:17:35just create a new project folder within
2:17:36your holding company,
2:17:38call it website or app or whatever and
2:17:40then just start getting Claude to build
2:17:41out in there.
2:17:42It's like
2:17:43>> Remi's not going to be signing with
2:17:44Lovable anytime soon. Most underrated AI
2:17:46tool right now.
2:17:48>> Composer.
2:17:49>> How long until a total beginner feels
2:17:51the productivity gain, like
2:17:52realistically?
2:17:54>> A few days. A few days.
2:17:57>> The task you'll never let an agent
2:17:58touch.
2:18:00>> Thinking
2:18:02and phone sales.
2:18:04box AI voice sales agents
2:18:07>> One MCP connection that changed the game
2:18:09for you.
2:18:11>> I'd say connecting Appify with all the
2:18:13thousands of online scrapers. You can
2:18:15pretty much get anything.
2:18:16>> Finish the sentence. In 2 years every
2:18:18founder will have a
2:18:22>> personal AI operating system.
2:18:24>> What is Obsidian and is it still
2:18:26relevant today?
2:18:27>> Obsidian is just a simple note-taking
2:18:30app that uses markdown, which is the
2:18:32love language of AI. The agents love it
2:18:34as we covered earlier.
2:18:35And a lot of people kind of build out
2:18:37there like I guess second brains or like
2:18:38that you know, the context folder that
2:18:40we uh built earlier. People just build
2:18:42like
2:18:43essentially a uh
2:18:45version of that context folder, but with
2:18:47hundreds or thousands of markdown files.
2:18:49And they just use Obsidian as kind of
2:18:51like the the layout of manage that, I
2:18:53guess. And I just think um I don't think
2:18:56it has its place. It's really good for
2:18:57like solo operators, but as soon as you
2:19:00try to like have a setup where you want
2:19:02your company company and team members to
2:19:03access to master those files too, it's
2:19:05messy. I don't think it's I don't think
2:19:07it's needed, especially for beginners.
2:19:09>> What is one skill that's changed the
2:19:11game for you?
2:19:12>> The Claude Council skill. So, I found
2:19:14this by a guy called Ali Lehman on
2:19:17Twitter and it's actually like my most
2:19:19viral video on Instagram talking about
2:19:20that skill. And you know, if you ask
2:19:23Claude for advice on something, if you
2:19:24say like, "Hey, why is this a good
2:19:25idea?" It's going to give you like a
2:19:26perfectly structured case as to why it
2:19:28is. And if you say like, "Why is this a
2:19:30bad idea?" It'll give you the same kind
2:19:31of case for why it's bad. And you know,
2:19:34it's why you can't trust Claude with
2:19:35decision-making. So, this skill
2:19:37essentially spins up a bunch of
2:19:39different Claude sessions, all given
2:19:42their own persona, like a contrarian
2:19:44that's going to hunt for what goes
2:19:45wrong, like an optimist, five different
2:19:47complete personas to basically like give
2:19:49their thoughts on your question. And
2:19:51then it gives it shuffles up their
2:19:52answers anonymously, gives it to five
2:19:54fresh sessions that are like the peer
2:19:57reviewing agents to like give their
2:19:58thoughts on those answers and then
2:20:00finally it gets into one new session
2:20:02which is like the chairman that like
2:20:04looks at all the final results and gives
2:20:06you the final verdict. I use it a lot
2:20:08for like decision-making cuz it's just
2:20:10interesting to get like an extra
2:20:11perspective on it. I'll use it all the
2:20:13time but I hey should we add this
2:20:15feature or is it too soon? Or it's like
2:20:17what should I work on next on the
2:20:19project?
2:20:20>> Sounds amazing. Before I get into last
2:20:21couple questions that I ask everybody,
2:20:22I'd love to just kind of recap and end
2:20:25with like
2:20:26where should someone out there just
2:20:28start? Yeah.
2:20:29>> So
2:20:31think of this entire system as like
2:20:33something it it it grows with you and it
2:20:36will continue to grow and compound. And
2:20:38you just need to start by planting the
2:20:39seed. Just get at your computer right
2:20:40now.
2:20:41Create that folder, the holding company
2:20:44folder OS.
2:20:45Start getting your context files built.
2:20:48Just open a fresh Claude or ChatGPT
2:20:50session and just say like help me build,
2:20:52you know, a bunch of markdown context
2:20:54files. We need like an about me, an
2:20:56ideal customer profile, a product
2:20:59services information, uh you know, brand
2:21:01voice and then interview me to extract
2:21:04all the answers from my brain and and
2:21:05build these documents. And then you add
2:21:07them into the folder
2:21:09and start Here's the big thing. Connect
2:21:12your tools as well. That's actually
2:21:13could even be the first step.
2:21:15Cool tools and context you can get done
2:21:17in a few hours.
2:21:19And then from there
2:21:22your goal is going to be to complete an
2:21:24entire work day from within inside
2:21:26Claude or whatever harness you're going
2:21:27to use. That's your first goal. And
2:21:30you know,
2:21:31it might be a little bit messy at the
2:21:32start but you're going to then start to
2:21:33build skills. Like if as soon as you go
2:21:35through a process with Claude like we
2:21:36did with the subject line skill, if you
2:21:39ask Claude to take a task off your plate
2:21:41and you know that you're going to have
2:21:42to ask that him to do that task again in
2:21:44in the near future or at least one or
2:21:46two more times, tell it to make a skill.
2:21:48And then that's the place to start. You
2:21:50just start growing this system and you
2:21:52can only do that through like working
2:21:54with it. So, get your your context files
2:21:56built, get that folder set up and your
2:21:58tools connected, and just start
2:21:59completing a work day from within inside
2:22:02your your Agent Harness.
2:22:03>> It's Challenger. What's up, guys? All
2:22:05right, three questions. Favorite book or
2:22:07podcast and why?
2:22:09>> My favorite book's The Alchemist. I
2:22:11don't know if you've read it.
2:22:12>> Well, Paulo Coelho. I'm Brazilian, bro.
2:22:14Of course I have.
2:22:15>> Dude, I love that book. I read it like
2:22:16four times. That's interesting.
2:22:18Yeah, something about
2:22:20>> podcast You a podcast guy?
2:22:21>> Yeah, I'm a podcast guy. I like the My
2:22:23First Million podcast.
2:22:24>> Good guys.
2:22:25>> And the Open Residency, yeah.
2:22:26>> Hey, shout out Sean and Sam Parr.
2:22:29>> They're guns, dude. They inspired me to
2:22:31start doing the newsletter. Like they
2:22:32they both crushed it with newsletters.
2:22:34>> Yeah, they're on the newsletter side.
2:22:36They're with Beehiiv as well, too.
2:22:37>> Yeah, yeah. They've They've They've both
2:22:38exited their newsletters for like tens
2:22:40of millions, right?
2:22:41>> Yep. Sam on The Hustle and then Sean on
2:22:43Milk Road, the crypto one. Yeah. That
2:22:45was a quick one, too. Entrepreneur or
2:22:47brand that you want to give flowers to
2:22:48and why?
2:22:49>> Brand?
2:22:52Composio. Dude, they're so good. Love
2:22:54Composio.
2:22:55>> never heard of Composio until today. We
2:22:57most definitely are going to put that on
2:22:59Stairs and put that on Everything, so
2:23:00I'm with you on that. What about a
2:23:01creator you want to give flowers to and
2:23:03why?
2:23:04>> Good question. Do you know Ask ChatGPT?
2:23:07>> Yeah.
2:23:08>> She's a beast, dude. Like I can spy me
2:23:10to start creating content and that Yeah,
2:23:11she's a she's a beast at
2:23:13content and AI. Big inspiration for me.
2:23:15>> Love it, man. I hope everybody enjoyed
2:23:17this episode. For the first time ever in
2:23:20Open Residency podcast history, I'm
2:23:22going to ask to pay it forward, guys.
2:23:23Share this with someone that you care
2:23:24about. Get people started in this game.
2:23:27This is definitely the most thorough. We
2:23:28took a long time to really kind of
2:23:30orchestrate an outline of what we're
2:23:31going to talk about, so I hope you guys
2:23:32enjoyed it. And uh we are going to put a
2:23:35[ __ ] ton of links and PDFs and tons of
2:23:38stuff down below in the description, and
2:23:40this definitely won't be the last time
2:23:41that we do stuff with Remi, so I
2:23:42appreciate all you guys. I appreciate
2:23:43you, man. It's been an amazing day.
2:23:45>> Thank you for having me on. I've I've
2:23:47enjoyed it. I think like this is just
2:23:49the crash course that everyone needs.
2:23:51Like if you're starting from zero,
2:23:53like this is just all you need. Just
2:23:55this episode and I think you're set.
2:23:57So, I'm glad we got it.
2:23:58>> Amazing. Thank you guys so much for
2:24:00watching. If you got value out of this
2:24:01episode, every [music] single principle,
2:24:03framework, tactic is in one free
2:24:05playbook, linked at the top of the
2:24:07description. Grab it, hit subscribe, and
2:24:09we'll see you in the next one.

About this transcript

This page contains the full transcript of AI Agents: The Most Valuable Skill You Can Learn in 2026 (Full Course) by Open Residency, generated from the public captions YouTube serves with the video. The transcript has 32,541 words across 4,932 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.