YT YouTube2Text

YouTube2Text → Transcripts → How I Plan, Build, and Run Loops with Claude Code in 40 Minutes | Thariq Shihipar

How I Plan, Build, and Run Loops with Claude Code in 40 Minutes | Thariq Shihipar — Transcript

by Peter Yang · 8,827 words · 1,218 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00As the models have gotten smarter, they
0:01need less direction, fewer constraints,
0:04and fewer examples. We've got /loop,
0:07we've got SL goal, and we've got
0:08workflows. These are all geared at
0:11trying to get the agent to run for long
0:13periods of time. This was just one shot,
0:14like when one prompt. So, it's created
0:16the caption, and it's created the little
0:18overlay, and now it's going to fade to
0:20black. My goal for this year is to be
0:22more productive, but work less.
0:27Hey, everyone. Today I'm really excited
0:28to welcome Tharic from the cloud code
0:30team. This has been a long time coming.
0:31So I'm going to ask Taric to show us how
0:34to design loops and workflows to get
0:35cloud to work longer and how he
0:37personally uses cloud code and also the
0:39new cloud tag. Welcome sir.
0:41>> Hey Peter. Yeah. Thanks for having me.
0:42I've been been excited to to join for a
0:44while.
0:45>> Yeah. All right. All right. So may maybe
0:46before you demo anything, maybe you can
0:48talk about at a high level how you and a
0:49clock team think about how do you go
0:51from you being a person prompting the
0:53agents to kind of design loops and some
0:55of the stuff to get the agent to kind of
0:57work by itself. Like how do you guys
0:58think about this?
0:59>> Totally. Yeah. Yeah. So I I think loops
1:02are a pretty general term I I think for
1:05just different ways of like having the
1:08agent get feedback or like sort of uh
1:11yeah work for long periods of time in in
1:13an orchestrated way. And so we've got
1:15slashloop, we've got slashgoal, and
1:17we've got workflows. And these are all,
1:19yeah, like you said, geared at trying to
1:21get the agent to run for long periods of
1:24time. Slashgoal is one of those things
1:26that it helps the agent remind itself
1:29what its exit condition is, right? And
1:31only allows it to exit once you know it.
1:33And so goal is really great for when you
1:37uh have a complicated task and you
1:39really need to, you know, make sure that
1:41this is like done at the end. you really
1:43want to prevent any stopping. You kind
1:45of also want to give the agent feedback,
1:47a signal to basically say like, hey,
1:49power through things, right? And and so
1:52I think one of the like if you were to
1:54step into Claude's shoes, sometimes you
1:55ask someone asked you to do a task, you
1:57run into a like maybe, you know, sort of
2:00a complication or something that is
2:02maybe not quite up to spec or like
2:04doesn't match what the user said, you
2:06might stop early and be like, hey,
2:07should I keep doing this or not? Right?
2:09And so uh goal is one way of just kind
2:12of the user indicating like hey I've
2:14I've done enough spec and expiration
2:17right I understand the problem space uh
2:19just go execute on it and if you run
2:21into something you know fill it in right
2:24so that's how we think about goal and
2:26and then workflows are in my opinion
2:28maybe like the most powerful form of
2:30this where you can you know spin up sub
2:32agents to both do the work paralyze the
2:34work and verify the work and and so uh I
2:37think especially for non non-technical
2:39work. This can be like a really great
2:41way of taking a non-deterministic task
2:43and then breaking it down roughly into a
2:45deterministic task. So yeah, I think
2:47that's how we think about some of these.
2:49And uh you know, I know I'm going to do
2:50a quick demo, but like before that, like
2:52just on the clock team,
2:53>> how do you guys use some of this stuff?
2:55Like do you do like a go make this match
2:58design or like go make this number go up
3:00or [laughter] what do you use? Yeah.
3:02>> Yeah. I mean, everyone does different
3:04things, right? based on like kind of the
3:07like um based on yeah what what they're
3:10trying to do. I I think like uh you know
3:13Jared's talked a lot about the the bond
3:15rewrite in Rust and how that use
3:17workflows and and he's going to be
3:19talking more about that. Um I think like
3:21anytime you have a deterministic sort of
3:24like signal like for example like
3:26latency, right? like um goal is like a
3:29great thing to just sort of like have
3:31the agent explore things sort of in like
3:34an auto research way, right? Um and uh
3:37yeah, I think that like
3:40you like let's say you're talking about
3:42design, right? I think one way you want
3:44to think about this is like how well can
3:46the agent like understand what your like
3:51design is to begin with, right? And and
3:53so um you might want to like turn it
3:56into like if it has a spec that it can
3:59verify like if it's a Figma file, you
4:01use the Figma MCP and then you're like
4:03slashgoal make sure that the rendered
4:06design matches the Figma MCP. That's a
4:08lot easier than like a screenshot,
4:11right? And and so if you're a
4:12screenshot, maybe you want to do a
4:14workflow uh where you like it's more
4:17squishy, you know, like there's like a
4:18rubric that you're evaluating against
4:21and you have a verification agent and
4:22things like that. So
4:24>> Got it.
4:24>> Yeah. Very dependent on the task.
4:26>> So basically the planning matters a lot,
4:28right? To give a because I tried to give
4:29it a go to like just build me an amazing
4:31game and [laughter]
4:33and then it kind of went off the rails
4:34because like you know just like one one
4:36line. So
4:37>> yeah, I mean I think um the the details
4:40of of like you know there's a lot to
4:44figure out in what you want, you know,
4:46there's a lot of work that goes into
4:47that. Yeah.
4:48>> Got it. Okay. So let's make this really
4:50concrete. Uh you've been sharing uh
4:52really awesome videos on Twitter about
4:55like uh using cloud to make the video
4:57and I think you have like a video
4:59workflow that you want to share with us,
5:01right?
5:02>> Yeah, exactly. So I yeah, I've had a few
5:04people uh ask about this and I think
5:06this is a good sort of demo both for
5:08like video editing but also just like
5:10how to think about like a nontechnical
5:12work. So 10 minutes before I got on this
5:14call, I recorded a quick video and so
5:18this is me just like you know making a
5:19video. Basically what I'm saying is like
5:21hey it's me um you know like I'm on the
5:24Peter Yang podcast. I'm going to point
5:27to where I want an overlay to show up
5:29and then I'm basically saying like,
5:31"Hey, uh, fade to black," you know? So,
5:33this is like a quick video that I that I
5:36put together, right? And, um, now like
5:40the prompt I gave it is this, right?
5:43Like, uh, this is a repo for Peter Yang
5:46podcast. Uh, there's a sample video here
5:48called Peter Yang recording. Transcribe
5:50it using whisper. Then use reotion to
5:53create a UI that shows the transcript
5:54with each individual word being
5:56highlighted and different overlays. And
5:58yeah, goal don't stop until the video is
6:00fully rendered. Right.
6:02>> And so now I like uh this was just one
6:06shot like when one prompt, right? And
6:08what it gave me is this. So it's it's
6:11transcribed it. It obviously doesn't
6:12know my name is Tharic, right? And uh uh
6:16it's like overlaid the the um the text
6:20uh or sorry it's created the caption and
6:22it's created the little overlay and now
6:23it's going to fade to black. Right. And
6:25so
6:25>> Oh, awesome. Okay. So So you basically
6:26said out the instructions out loud,
6:28>> right?
6:29>> I said it out loud. Yeah. Exactly. And
6:30and of course there are many ways that
6:31you can do this but it's like now in
6:34this directory you know uh transcribed
6:36everything right. Yeah. Uh using whisper
6:39uh it's created the reotion stuff. Yeah.
6:42This is like the very basic start to uh
6:44video editing workflow that I have which
6:46is uh yeah I think a good signal for
6:48like how to do non-technical work in uh
6:51in cloud code.
6:52>> So you kind of did this manually with
6:53the prompts stuff but then you probably
6:55saved it into a skill or something
6:56right? Yeah.
6:57>> So right now I don't have a skill. I I
7:00think one of the things I try and do
7:02first is really figure out what I want
7:03before I turn it into a skill you know
7:06and so what are the like uh edge cases
7:09here? Um, one of the things I noticed
7:12that I had to prompt it towards is is
7:14getting better at figuring out where my
7:17um where my hand was, right? And so like
7:20or where I was pointing towards. And
7:22even now I'm not not thrilled with like
7:25where this showed up, you know? And so
7:28[snorts] like yeah, like you could
7:29imagine that one direction I could take
7:31this in is I could be like, oh, I want
7:33to track my fingers or track my face and
7:35then get give the agent more metadata,
7:38right? so that it can then do
7:40interesting overlays.
7:41>> This episode is brought to you by
7:43Whisper Flow. Whisper Flow saves me at
7:45least 3 hours a week and is one of my
7:47favorite AI apps by far. It's just so
7:49much faster to dictate to AI using your
7:51voice than to type. You just talk
7:53naturally and it outputs clean, ready to
7:56send text. Whisper Flow even removes
7:58filler words and formats your sentences
8:00for you. I use Whisper Flow for
8:01everything, including drafting
8:03newsletter posts, writing product specs,
8:05replying on Slack, and more. It works on
8:07Mac, Windows, iPhone, and Android across
8:10all of your favorite apps. Try a free at
8:12whisperflow.com and use my code
8:14peterwisperflow to get 6 months free.
8:17That's Peter Whisperflow. Now, back to
8:19our episode. I think you had another tab
8:21that had like a plan or something. Is
8:23that for the video skill?
8:24>> Yeah. Yeah. So, I I think I want to talk
8:27uh I'd love to talk a little bit about
8:28how I got here. You know what I mean? So
8:31like I think like even to do this prompt
8:34I think it took a bunch of uh planning
8:37and learning to get here right and then
8:39I think that like um one of the things
8:43that I I I think when we talk about
8:46plans we often are talking about things
8:48where it's like um just one shot right
8:51like you plan and then you do something
8:55and that that's it. But I think planning
8:57is more of this iterative process of
8:58like exploring, investigating, finding
9:02out what you don't know, what you want,
9:05right? And then like together like in
9:08the end that sort of simplifies itself.
9:10But so like here's an example. But while
9:13we do this, I I would love to do
9:15something where it's like, okay, um I'm
9:18going to give it okay to say like I want
9:21to update the UI of these overlays,
9:25right? Um and I want to use
9:30I want to use Peter Yang's style. Um
9:35here's his blog.
9:38uh create a HTML artifact for exploring
9:44different designs for the overlays and
9:48captions
9:49uh design variations.
9:52Okay, so this is like an example to me.
9:55This is also planning, right? Because
9:57like what I'm doing is sort of like I um
10:00I'm trying to figure out what I want
10:02here, right? And and so it's more of
10:03like a exploration. You can see one of
10:06the important things I'm doing is like
10:07I'm giving it a reference, right? So
10:09when I give it, you know, your website,
10:11it can now fi fetch the HTML and start
10:14doing this. So like this is sort of like
10:16a plan that will will run while I sort
10:18of go over the other, you know, kind of
10:20things I do when I'm planning, right? So
10:22one of them is like okay, what how does
10:24transcription work, right? And and this
10:26is actually I think pretty important to
10:28know because like it informs a lot of
10:30the edge cases. So uh this this
10:33transcription uses whisper and so um uh
10:37you know like cloud code sort of put
10:38this explainer together which is
10:40honestly like kind of amazing. Um the
10:43important thing for me is like the uh
10:47like ways things can go wrong, right?
10:48And this is something I prompted it in
10:50the plan was like I want to explain
10:52whisper to me and understand what the
10:53edge cases are, right? And so one of the
10:56things that says like silence can become
10:57thanks for watching, right? Um or yeah,
11:00a word can be split into two two chunks.
11:02It doesn't know it doesn't have speaker
11:04recognition, right? And so there are a
11:07bunch of different like edge cases here
11:08where um I I think all of this is quite
11:12like good and this sort of like helped
11:14me build a confidence in using whisper,
11:17right? Yeah.
11:17>> Um but having these edge cases and
11:20knowing sort of what the limits are
11:21upfront really helped me like avoid this
11:25case where I I build this complicated
11:27workflow during whisper and then you
11:29know I realized there are things going
11:30wrong um and I like sort of didn't have
11:33these unknown unknowns, right? And so
11:36this case is sort of like me sort of
11:38doing planning, but really it's like me
11:40discovering my unknowns. And um I think
11:43that can take a lot of different shapes.
11:45It can be learning, it can be technical
11:47specs, right? It can be mock-ups and
11:50exploration. Um yeah, I I'm starting to
11:53like I feel like the word planned is
11:55maybe uh too broad right now. So
11:57>> is it's it's more like exploration and
11:59understanding like what you're trying to
12:00do basically.
12:02>> Yeah. I like to say uh like getting rid
12:04of your unknowns, you know, like I think
12:06whenever you have a task,
12:08>> it it kind of like almost always uh
12:11there's a lot that you don't know either
12:15you don't know how things work or what
12:16you want and it's very very iterative,
12:19you know, and so it's not just like a
12:20you write it all down once and then you
12:22implement it. I think there's like many
12:24steps and uh different passes. So
12:26>> and yeah, and and dude, this is like a
12:28beautiful plan, man. Like, dude, do you
12:30have like is this like the claw template
12:31or something? Like it has images.
12:33>> This is using the front-end design
12:34plugin. I think you know like I think
12:36like uh yeah, it's not uh yeah, there
12:39are uglier plans here. Like I've gotten
12:40this like reotion plan. Um
12:44>> I I think that like you know the the
12:45design of it doesn't matter so much as
12:47sort of like you really just want to
12:49make sure it's something that you really
12:52do read, you know, and kind of get a
12:54sense of like not all of it, but like
12:56there's important parts in it. And I
12:58think one failure mode I see is that
13:00people still sort of like glaze over,
13:02you know, the plans and explainers. So,
13:05um, yeah.
13:05>> Yeah, it becomes really because like AI
13:07can write all these, you know, crazy
13:08markdown files and they're usually
13:09pretty long and then at some point like
13:11I just get lazy. I'm like, okay, you
13:13know, just just just do it. Just
13:14[laughter] do it.
13:14>> Yeah. Yeah. Yeah. Exactly. I I think
13:16that's like the thing that, you know, it
13:17happens to all of us, right? Like a like
13:20the prompt box can definitely just be a
13:22lazy button, right? Where you're just
13:23like, hey, just do the thing. Um, but
13:26usually, you know, like uh you end up
13:29paying for that, right? Because if
13:31you're if you're trying to do something
13:32serious uh and you know, like you're
13:35taking the lazy step at each way, it
13:38it'll it'll end up taking longer, maybe
13:39costing longer, too. So, um,
13:42>> so it's kind of this iterative. So, so
13:43you ask about like the the whisper and
13:45then you kind of like learn a little bit
13:47more and then you ask to make another
13:48plan and then at some point you have
13:50like a plan that you can share, right?
13:51>> Yeah.
13:52>> Exactly. Yeah. Yeah. So I I I I did a
13:54lot of for this video recognition stuff.
13:56I did a lot of like you know research
13:59into how different video like algorithms
14:02worked, right? Like I I think I did one
14:04on like video segmentation. Um I wanted
14:07to put like text behind a subject and so
14:09this was like something I like explored
14:11and learned more about
14:12>> and um I ended up finding that like
14:14there wasn't something reliable enough
14:16here for me to use. But um yeah, I think
14:19there's like just a lot of this like
14:21iterative process of like finding out
14:22what do I want, you know, and what is
14:24possible right before I can do it. And
14:26so um it looks like this overlay is
14:28maybe almost done. So we'll we'll we'll
14:30see it soon. Yeah.
14:32>> And how about like on the team? So have
14:34you got the rest of the team to review
14:36HTML? Is that a markdown or
14:38>> Yeah. Yeah, for sure. I mean I think
14:40like everyone is different, right? So I
14:42think that like but I think HTML
14:44artifacts I mean we launched artifacts,
14:46right? Um, currently it's only on teams
14:48and enterprise, but um, hopefully coming
14:50to Maxin Pro. Um, and uh, yeah, that
14:54that's like how we share things now at
14:56the company. It's like we ask Claude to
14:58make an artifact and this can be, you
15:00know, of a plan of a PR that we've
15:02already done. Uh, it can be, you know,
15:05like a status report, like an incident
15:07report, things like that. So, um,
15:10>> yeah, it's, uh, definitely how we do it.
15:12Yeah.
15:13>> All right, let's see what it generated
15:14here.
15:15>> Yeah, let's let's do it. So let's say
15:16open overlay style HTML. Okay. So yeah,
15:21here are some of the like uh you know
15:24the options it's made. I think it's
15:26probably maybe like indexed more on the
15:28stack over or the uh Substack brand
15:31versus your brand. You know what I mean?
15:32>> Yeah. Because I don't I don't have a
15:34brand. Yeah.
15:35>> Yeah. Yeah. Yeah. No, I think you do,
15:36right? I think you do. Like I think
15:37you've got the red and white. We can
15:39probably iterate on it there. Um but
15:41this is one of the things where I'm like
15:43you know there's quite a lot of
15:44difference between these different uh
15:46like you know designs right and um yeah
15:50one of the ways I like to to plan is is
15:53to like do this exploration especially
15:55because I'm not a designer um so like I
15:58just really just only know it when I see
16:00it you know.
16:01>> Got it. Got it. Um like just to wrap up
16:03the plan conversation like
16:05>> yeah you know I've been writing like
16:06product specs for like 10 years and like
16:08you know they usually have like what's
16:09the problem we're solving what's the
16:10solution what's the goal so on and so
16:12forth right but I feel like part of this
16:14stuff is like read by agents now
16:16>> and I I I feel like the different
16:18sections of your spec needs to change or
16:20or maybe like the product spec and
16:21technical spec is like one thing like
16:22how how do you guys do it on the team
16:24like you have like one section for the
16:25humans to read and one section for the
16:26agents to read or or how do you think
16:28about it?
16:29>> Yeah, good question. Good question. I I
16:31think that it is very tied together,
16:32right? Like I I think a spec kind of can
16:35evolve even towards like planning like I
16:38think one of the things I think I'm
16:40showing here, let me see um
16:44implementation.
16:46Yeah, I think like one of the things
16:47that you sort of see here uh sometimes
16:50is that when you run something the uh
16:55the model can find things that it didn't
16:57anticipate or you didn't anticipate when
16:59you're implementing, right? So I see
17:01>> um this is like I I don't think of
17:04specking just happening at the start,
17:06right? It's like you you start with like
17:08the human request and then the agent
17:11does some like technical exploration. It
17:13comes back, maybe you do some mockups,
17:15some explainers trying to understand
17:17your unknowns. You refine that. Uh you
17:20give it to the agent again. It might
17:21start implementing it. uh I ask it to
17:23keep implementation notes as it goes so
17:25that it like finds out you know like
17:28what are the you know what are things
17:30that we we're not expecting about this
17:32implementation and once we have that we
17:35can actually you know respspec if we
17:37need to right like it depending on how
17:38things go. So yeah I think it's much
17:41less like there's like one handoff of
17:43spec to implementation and more this
17:46like back and forth process. Yeah.
17:48>> Yeah. cuz I guess it's pretty cheap to
17:49build now. So you can just like maybe
17:51ask it to build the simplest version of
17:52this and like there's probably a lot of
17:54bugs and issues and you can keep
17:55iterating, right?
17:56>> Yeah, exactly. Or like the prototype
17:58version of this, right? Like that's what
18:00how I think about these overlays. These
18:01are like prototypes of the, you know, of
18:04the design, right? And
18:06>> okay,
18:06>> um if we like uh
18:09>> like it then we can do the more
18:10expensive version, which is instead of
18:12HTML, we can do it in React. that means
18:14you have to like you know sort of
18:15rerender the video and do all the code
18:17changes. Um so like what's the like
18:20smallest step you can take to prove out
18:23the concept that you want to like you
18:26prove out the spec more. Yeah.
18:27>> And like I Yeah. I love how it takes
18:29implementation. So So the next time you
18:30start a new section, it can just refer
18:32to the HTML, right? It's like a living
18:34document.
18:35>> Yeah. Exactly. Yeah. Yeah. Got it. What
18:38percent of your work now is done through
18:40clot tech on Slack versus in like the
18:42terminal or like you know the cloud app?
18:45>> Yeah, I I think that like the way for me
18:48it's trending is that like there's a lot
18:50of parallel stuff happening in cloud tag
18:52and so any multiclouding is often
18:54happening in cloud tag unless there's a
18:56reason why it needs to happen locally,
18:58you know. Um I think one of the things
19:00we sp put a lot of time into is like
19:02making sure that our environment can run
19:04um in in remote and so um yeah I I
19:09multi-clotting is happening in cloud tag
19:11you know initial explorations PR specs
19:13like trying to understand something is
19:16happening in cloud tag and then once I
19:18get into there's usually one thing I'm
19:20focusing on and that's happening in
19:22cloud code and it's like more you know
19:24like back and forth iterative right um
19:27and So yeah, I think that's how I'm
19:30using it, but I think everyone at the
19:31company has sort of like different like
19:33approaches and mixes to it.
19:35>> And when you say like what do you mean
19:36by multicloud? Oh yeah, like like just
19:39like if I have like multiple tasks
19:40happening at once, I especially if I'm
19:42like trying to involve someone else,
19:44like if I am um you know, I've got a PR
19:48that I'm trying to get merged, you know,
19:49like what one thing I'll do is like I'll
19:51you know, ask it to babysit uh the PR
19:54fix test and then tag a reviewer and my
19:57like reviewer will get tagged in the
19:59same slack channel and we'll like be
20:00able to talk there. Um but just
20:03generally multiling is like anything any
20:05background work where before I was
20:06having you know like five different
20:08clouds uh in my cloud code now like it's
20:11mostly one active cloud code session and
20:13then a bunch of cloud tag sessions.
20:15[snorts]
20:16>> Oh interesting. Okay. And um so you talk
20:18to cloud tag both through like DMs on
20:21Slack and also like in share team
20:22channels I'm guessing or
20:24>> um
20:25>> primary. Yeah.
20:26>> Yeah. In in Slack. Yeah. Yeah. and both
20:28in my private like like I have a yeah my
20:30thoric claude slack channel where I do
20:33like most of my work and then uh yeah we
20:35have like team channels like feedback
20:37and things like that where we um or like
20:40project specific engineering channels as
20:41well um that's like a common pattern we
20:44see yeah I guess that makes sense
20:45because I I do think like uh these
20:47coding apps right now are like primary
20:48single player experiences like you're
20:50kind of talking to the agents through
20:51different threads
20:52>> and and if you think about it like the
20:54the cloud app actually kind of looks
20:55similar to Slack's like a bunch of
20:57threads Like just like Slack has a bunch
20:58of channels. So I guess like the idea is
21:00like Slack is now the multiplayer cloud
21:03experience, right? Because like
21:04everyone's there already.
21:05>> Yeah. I mean I I think like this is you
21:07know where it's starting, right? Like I
21:08I think ultimately we think cloud will
21:10be you know this like sort of proactive
21:11agent that's sort of like meeting you
21:13where you are you know and I think Slack
21:15is where anthropic is right. Um and uh
21:18yeah it's a very natural way to explore
21:20and yeah it's surprisingly good at
21:21coding. can like uh you know there are
21:24some people who just do all of their
21:25coding almost completely in cloud tag.
21:27Yeah.
21:27>> But okay maybe this is a dumb question
21:28but how do you like uh like you know in
21:30cloud code like easily trigger scales
21:32and stuff but like if I'm in cloud tag
21:34like do I just tag cloud and slash the
21:36scale or how is it?
21:38>> Um yeah you can just tell it to use the
21:40scale.
21:40>> Okay. Okay. [laughter]
21:42>> Yeah. Yeah. Yeah. There is still I I
21:44think there's still UX iteration we're
21:46working on around all of this stuff.
21:48Yeah.
21:48>> Okay. I guess do you think uh do you
21:50agree like I feel like the future for
21:52this like because you you gave me a talk
21:53right on on human agent interaction
21:55>> and like um I feel like the future is
21:57just like the agent is just like another
21:59employee like you got you got to onboard
22:01an employee you can talk to it through
22:03Slack you can like give give it a phone
22:05phone call you know we'll have to think
22:07>> yeah um I think it's a good question I I
22:10think that
22:12um I think these metaphors can sometimes
22:14be helpful but also limiting in some
22:16ways I I think that like One of the
22:19things you know when we talk about
22:21identity is that like in cloud tag every
22:23agent every channel has its own memory
22:26right and so um that is like I think one
22:30choice right like you could also imagine
22:31that there are like uh you know multiple
22:34clouds you tag like each one has a
22:36different slack identity and you're
22:38tagging them and things like that and
22:39and um I'm not like I I think it is kind
22:43of helpful to think of an agent that is
22:45like sort of has persistence and uh you
22:49know uh memory but um I think they're
22:53also different from co-workers in in in
22:55some ways right but I think they're like
22:57they're proactive and like you know yeah
23:00it's proactive it has memory it has
23:01identity um but yeah it's like uh just
23:05sort of like an evolution of cloud code
23:07and I think we want to see where like
23:09you know uh where the models will uh
23:13will take us there versus sort of like
23:14putting it into a box you know
23:16>> yeah that's a good point yeah Like it
23:17definitely has a lot better memory than
23:19any human. [laughter]
23:21>> Better and worse sometimes. You know
23:23what I mean? Like it's spiky, right? So
23:25yeah.
23:25>> Yeah, makes sense. Okay. All right,
23:27dude. Why don't I do something fun? I I
23:29want to show you my clock setup. I think
23:30we talked about this in person. I feel
23:32pretty proud of this. I I built this
23:33podcast production skill
23:35>> and uh what it does is like it takes a
23:38transcript of an interview that I did. I
23:40I actually interviewed your colleague
23:41Jess and uh it it kind of generates a
23:44bunch of stuff for me. So it generates
23:46thumbnails and like stuff to cut. So
23:49basically I I just tag it and I paste a
23:51transcript and then it starts generating
23:53like you know clickbaity YouTube
23:54thumbnails and stuff. [laughter]
23:56>> Yeah.
23:57>> Yeah. So it generates stuff and and like
23:58I I basically give it like my examples
24:01to try to keep it on track.
24:03>> Yeah.
24:04>> But I guess that this skill is kind of
24:05like trying to do a lot of things. So
24:06maybe like I'm curious if you have any
24:08feedback or you know.
24:09>> Yeah. This is cool. Can I see what it
24:10produces like the the clips or like you
24:13know
24:13>> Oh what it produces? Yeah. So some of it
24:15what produces is is just like pure text.
24:17It's like here's a news post here some
24:18takeway
24:19>> copy into YouTube. Yeah.
24:20>> Yeah. And then I have this other skill
24:21called the video post skills. It's kind
24:23of similar to yours, but I'm sure it's
24:25not as sophisticated, but like it
24:27basically takes the YouTube video for
24:29this thing.
24:30>> Cool.
24:30>> And then it uh extracts it and and then
24:33it gives me some ideas for clips to to
24:36make.
24:36>> Yeah.
24:36>> Yeah.
24:37>> Does it do the clips as well?
24:38>> Yes, it actually does. So I I said do
24:40two and then it uses uh I don't actually
24:43know what it uses. It uses a bunch of
24:44random stuff to to
24:45>> probably fmp and stuff, right? Yeah.
24:47>> Yeah. Probably fm and then it makes a it
24:49makes a video. Yeah. And it adds
24:51captions.
24:51>> Yeah. Yeah. That's awesome.
24:53>> So So yeah, it's not perfect though, but
24:55but yeah. Mhm.
24:57>> Yeah. I mean, um this is great. How how
25:00like
25:02uh how well do you feel like it
25:05matches what you want, you know, like or
25:08do you know like what it would look like
25:09to be better? Um, I wish it could like
25:12add some like, you know, B-rolls and the
25:14stuff that you showed me like overlays
25:16>> and stuff like that
25:17>> and like I I I wish it was smart enough
25:18to know how to pull in like for for the
25:21video clip like um, you know, if the
25:22cloud logo or or something you can just
25:24find it online.
25:26>> Yeah.
25:27>> Yeah.
25:27>> So, should I should I just tell cloud
25:28code to do all this stuff or
25:29>> I I think you can. Yeah. I mean, I think
25:31that like are are you working in one
25:33persistent repo or like um I think one
25:36of the things I think about is like
25:38sometimes you want a skill sometimes you
25:39want a repo with a lot of scripts and
25:41things that you are like you know like
25:43it's more of like a a workspace that
25:46you're working in, right? And like I
25:48think of a skill Yeah. Like I I think
25:50sometimes a skill can be instructions on
25:51how to create that workspace, right? And
25:53so
25:54>> um that that's maybe something to think
25:56about because like the more you know
25:58scripts and sort of like things that you
25:59build up the more the agent can the less
26:02it has to do from scratch you know um
26:05but uh yeah
26:07>> I mean I folder each time
26:09>> no I have all my skills at the user
26:11level
26:12>> uh and and there's like this personal
26:15whatever personal OS folder that has
26:17like all the output and stuff.
26:18>> I see. I see. Okay cool. Yeah. Yeah.
26:20>> But but you're saying like
26:22>> Yeah. I mean, I think you can sort of
26:23build like a video editing harness here.
26:24I I'm I'm uh curious about the
26:27thumbnails. Do you ever use like an
26:30image gen API like a Gemini or even an
26:32OpenAI or something like that? Yeah,
26:34>> I found it really bad at uh changing my
26:37face. So, like for example, if I have a
26:38smiling thing and that has to make it
26:40like a shocked face,
26:42>> it makes me look super ugly. Dude,
26:44[laughter]
26:46>> that's funny. That's funny.
26:47>> But um
26:48>> Yeah. Yeah. But I think it's pretty good
26:49at like like I don't if I give up my
26:51face and I tell you to change the
26:52background of the text, it's pretty
26:53pretty good.
26:54>> Yes.
26:55>> Yeah.
26:55>> Yeah. I mean I think Claude is quite
26:57good at like like one of the things that
26:59Cloud is good at is using other tools.
27:01So like you could give it like the
27:02Gemini or OpenAI image gen APIs
27:05>> and then you can ask it to look at your
27:08faces that it generates as well, you
27:10know, and sort of like tweak it there.
27:12So it can do this like interactively and
27:14pro progressively as well which is which
27:16can be helpful you know.
27:18>> Oh okay interesting.
27:19>> Yeah.
27:19>> So so basically so basically what I do
27:21is I I basically just chain a bunch of
27:22skills to together like like you know
27:24>> cool
27:24>> prepare this thing and then use a
27:25thumbnail and so on and so forth.
27:27>> I feel like I I know how to use skills
27:28now but like some of the other stuff you
27:29talked about online like dynamic
27:31workflows and stuff
27:32>> like I have no clue. Right. I have no
27:34clue.
27:34>> Yeah. How would we use a workflow for
27:36for something like this? Right. I I
27:38think like one thing you can imagine is
27:40like when you're using
27:41>> I I think your thumbnails example was a
27:44good one where or sorry your shorts
27:46example was a good one uh where maybe
27:48you want to generate 10 different shorts
27:51um or five different shorts or
27:52something. Workflows are a great example
27:54for that, right? So you can like say,
27:57hey, like um the main agent decides
28:00which five areas of the clip, you know,
28:03you want to like create a short for,
28:05right? And then the like workflow spins
28:09off a sub agent for each one. You also
28:11give it maybe a rubric for like what
28:13does a good clip look like, right? Like
28:15uh and then each one is verifying
28:17against it. Um, and then you're getting
28:20back at the end this like um, like you
28:24know, each clip has a maximum amount of
28:26compute put into it to make sure that
28:28it's matching what you want versus like
28:30if you're doing two or three
28:31simultaneously sometimes like I I think
28:33like um like Claude might like verify or
28:37put less work into any individual clip.
28:40So um,
28:41>> I see.
28:42>> And and to create a workflow, do I just
28:44like uh just tell to make a workflow
28:46basically? Yeah. or just be like, "Hey,
28:47like yeah, you know, create 10 clips
28:50here. Use a workflow. Uh this is my like
28:53rubric to to verify uh like what a good
28:57clip looks like, right? And this can
28:59also be a skill. So you can package a
29:01skill with the workflow. Um and then
29:04like you can the workflow is just a JS
29:06file and you can like ask it to like you
29:09save the JS file into the skill and now
29:12you have this like reusable skill that
29:14you can use. And this is like a dumb
29:15question, but like the main advantage
29:17over doing this workflow thing versus
29:18just using a scale is like uh it can sp
29:20sub agents and like keep the context
29:22clean or um yeah. So I I think it's like
29:26yeah, one of it is context. I think the
29:29other one is sort of like um laziness
29:33and verification, right? So like uh if
29:36you have a you know like a a short is
29:39one example of something where um you
29:42don't have a deterministic like hey is
29:44this a good short or not right and so if
29:46you have a rubric and you have a uh like
29:51uh you know verification agent that
29:53reads the rubric and then it looks like
29:55oh like hey let me review the short to
29:58make sure it's good and give feedback
30:00right uh that's something that the
30:02workflow enables as well and and we
30:03found that like you know we call it
30:05self-referential bias like when a model
30:07prefers its own outputs it's going to be
30:10like more lenient at like um uh it's
30:15going to be more lenient at like you
30:16know verifying it. Okay. So basically
30:18you want to have a separate cloud doing
30:20the work versus verifying the work,
30:22right?
30:23>> Yeah. And in this case like a separate
30:24cloud coordinating the work uh so that
30:27your main agent is coordinating it
30:29coordinating it each one doing the work
30:31and and then verifying. Yeah.
30:33>> Okay. And this way like the three
30:34separate clouds have like different
30:35context models and it's not just like
30:37less bias.
30:38>> Yeah. Exactly. And and they will both
30:40all use more compute I think like
30:41they'll put more like you know less
30:43likely for it to like stop early. Yeah.
30:45Think more. Yeah. Okay, got it. Um,
30:49awesome, dude. Well, let me ask you some
30:50like high level questions, dude. Since u
30:52it looks like you do you mostly use
30:54cloud in your in the terminal or do you
30:56actually use the app?
30:58>> I'm a mix of terminal and desktop. Yeah,
31:00depending obviously we do a lot of like
31:02what we call ant fooding. So, it's sort
31:04of like um you know, whatever I feel
31:06like I need to test the most. Yeah.
31:08>> Got it. Do you have any tips for like I
31:12I I feel like sometimes I get exhausted,
31:14man. Like like there's like five threads
31:16going on. at once for five different
31:17things and then like I'm I'm just like
31:18they're constantly paying me. It's
31:20actually worse than having backto-back
31:21meetings in some way. [laughter]
31:23>> Any tips for how to keep your own cult
31:25window clean
31:26>> or you know?
31:27>> Yeah, I mean this is a good question you
31:30know I think that this is like my goal
31:32for this year is to be more productive
31:34but work less you know and so um I think
31:38this is like I think something we should
31:40all sort of be able to push ourselves
31:41on. Um I think that like
31:45uh you know like I I think one thing we
31:49try and do is sort of like I try to have
31:50like one project I'm focusing on you
31:52know and and so even if there are other
31:54things that I'm just need to get
31:55unblocked like I need to have this build
31:57and merge and you know things like and
32:00explore something like I think one
32:01project that I'm really focusing on is
32:03very um helpful just uh because I find
32:06that the thing that cost me the most
32:08time is when I'm like yeah maybe a
32:10little bit lazy and and now I've like on
32:12a bunch of I'm like sort of multitasking
32:14these things. I do a lazy prompt and
32:16then I'm like, "Oh, like I've wasted
32:18this time now," you know, and so um
32:21yeah, I think that there's like just
32:23because like I think there's a optimal
32:26amount of multi multiclotting depending
32:28on, you know, who you are, what you're
32:30doing. Um but for me at least, it's
32:32something where like I tend to have like
32:34one task where I'm really focusing on
32:35the most. Yeah.
32:36>> Okay. So maybe like the AA agents take
32:38care of the rest or something, but like
32:40why not?
32:40>> Yeah. or even prioritization for
32:42yourself, you know what I mean? I I
32:44think can be like a good like angle, you
32:45know? Yeah,
32:46>> makes sense. And and and like um do you
32:48have concerns? Like I I worry that um
32:51like you know when because the agents
32:53like work pretty hard and keep going and
32:54and like
32:55>> Yeah.
32:56>> Maybe this is my problem like I I don't
32:57actually like like I said, I don't
32:59actually read everything that it
33:00produces carefully. Oh, [laughter]
33:02so then so then I worry that like the
33:03whole repo will just turn to slop at
33:04some point like if I don't
33:06>> like do you have any like uh routines or
33:08jobs that clean this stuff up
33:09periodically or
33:11>> Yeah. So I mean um obviously we have
33:15things like simplify which Boris put out
33:17that like you know sort of simplifies
33:19the repo. I think it depends on what
33:20you're using it for as well, you know,
33:22like I I think someone like uh when
33:25you're using it for the outputs, like
33:26you're using it for the like the video
33:30outputs. um those are pretty uh like you
33:35know maybe the quality of the code
33:37matters less you know and the agents are
33:40very like you know persistent and and
33:42they'll figure it out but um I find
33:44organization is often more for me than
33:46the agent to sort of like you know make
33:48me feel better about the workspace um if
33:50I'm only caring about the outputs but
33:52yeah you can always just ask it to to
33:53simplify or organize um yeah
33:56>> okay okay got it and and you try to keep
33:59this con do you try to like minimize uh
34:01and try to keep the concept window clean
34:02like you try not to like have a bunch of
34:04you know MCPs on or like you know cloud
34:07MD has a super long like do you try to
34:09optimize that or
34:11>> yeah I mean I I think that like yeah I
34:13yes I I think that one of the things
34:16we've noticed especially as the models
34:18have gotten smarter um is that like uh
34:22you know we'll be talking more about
34:23this we we cut down the cloud code
34:24system prompt by 80%
34:27um and the reason is that as the models
34:30have gotten smarter, they need less
34:32direction. They need fewer constraints
34:34and they need um fewer examples, you
34:38know, and so a lot of our system prompt
34:39would be like, okay, like here's the
34:41batch tool. Here are like five examples
34:42of using the batch tool. Never do it in
34:45these cases, right? And um the models
34:48are now kind of aligned enough that they
34:49know like hey uh we don't need to um
34:55like uh we like we don't the examples
34:58are almost constraining it you know
35:00because now it's like uh oh like you
35:02want things like this example and so if
35:05if you remove the examples it actually
35:07can be more free format and like the
35:09constraints can also constrain it too
35:11because like uh you oftentimes when you
35:15say never you really mean never. You
35:16just mean like most of the time don't do
35:18this. And if you give it the reason you
35:21don't want to do it instead, that can be
35:23more effective than the like don't do
35:25this constraint. So, um I think that
35:28like all this to say is that like you
35:30want to trim your context. I feel like
35:32cloud. MBS are probably too long right
35:34now and you probably want to shorten
35:36them more and more, right? Uh probably a
35:39lot of skills are too long. Um I think
35:41MCPs, you know, depends on the MCP. I
35:44think like obviously some of them can
35:46take up a lot of context but I think a
35:47lot of the MCP team has put a lot of
35:49work into this to you know make it
35:51better with tool search and things like
35:52that. So but yeah definitely cloudds
35:54like any sort of instructions the models
35:56just need like more room to run oftent
35:59times. Yeah. Then do you like instead of
36:01saying like uh you know like for example
36:02let's say writing a Twitter post like
36:04instead of saying like hey make sure
36:05it's like
36:06>> 280 characters or less and do not do
36:08this like do you give it like more like
36:10principles to follow like you know like
36:13how do you prompt it then like is it
36:14more like
36:15>> yeah I think a Twitter post is a good
36:17example where you might give your give
36:19it context on yourself. So you're like
36:21hey I am you know I work on cloud code
36:23I'm anthropic uh this is sort of like
36:26some of yeah some of the principles we
36:28follow. Obviously 280 characters is
36:30still like a good like um sort of
36:33constraint because that's like it does
36:34need to happen, right? But you can even
36:36just say it's a tweet and it will like
36:37sort of know um like or this is a good
36:40example where you might say like hey
36:42keep it under 280 characters. Uh but
36:45let's say that maybe there's a version
36:47of it that's better as a tweet thread
36:48with two tweets. You know what I mean?
36:51And so like in giving it the sense of
36:52like hey I want to write a tweet thread.
36:54I'd prefer it to be one tweet, you know?
36:57um gives us more freedom and
36:59flexibility. Yeah. To to find something
37:01good.
37:01>> Okay. Got it. Okay. Um and and then let
37:04me ask this. Um so Boris has been saying
37:06like uh coding is like a solved problem,
37:10right? And like
37:12>> Yeah.
37:12>> Um and and I feel like there's like
37:14becoming more technical these days is
37:16like a little bit like I guess my
37:17question is for someone like me who
37:19wants to actually, you know, learn how
37:20to work with agents more to actually
37:22learn what they're doing. Like how do
37:24you actually become more technical? like
37:25like it's not about the syntax or
37:27anything, right? Is that is it just like
37:28using it more or like how do you
37:30>> Yeah, I mean I think for the first step
37:32is motivating yourself to learn things,
37:34you know what I mean? I I think that is
37:36like um honestly really hard and and
37:40like I find myself doing this as well
37:42where like uh you know if you don't need
37:45to learn something to get the job done
37:47maybe you won't right but I think that
37:49like um it is really important. I think
37:52the goal of learning to be more
37:54technical is to know my unknown
37:57unknowns, right? And so there are some
37:58things that like being technical and
38:00knowing the syntax of Typescript is not
38:03really very helpful, you know? But I
38:05think like being technical and knowing
38:07like okay like hey uh what are the like
38:10trade-offs of different backend services
38:12and and uh what are the different video
38:14encryption libraries and how do they
38:16work and what's the difference between a
38:17local or remote video encryption or
38:19transcription library right these are
38:22all like I think pretty helpful right
38:24and so I think oftentimes I'm trying to
38:26learn the constraints of the system
38:28right and so like what is possible how
38:30is it doing it right now how good could
38:32it be you know like what if we did
38:34something else. And um Claude can often
38:37brainstorm and teach you this if you
38:39like push it, right? But you really do
38:41have to push it. I think like um this is
38:44like, you know, just the hardest thing
38:46about education and what everyone says
38:47is sort of like it feels good to like
38:50try and learn something, but like
38:51actually learning something is more
38:53work, you know, and like it should feel
38:55like work. I think Kaparthi says this a
38:58lot, right? Like education should feel
38:59like work more than fun. So yeah.
39:02>> Yeah, that's that's actually good. cuz
39:04cuz um cuz actually easier to just like
39:06for example with the videos thing,
39:07right? Like it's actually easier to just
39:09keep prompting cloud to just do it and
39:10then look at the output and then like
39:11like not actually learn anything, dude.
39:13Like just like go figure out and then it
39:15figs out. But the fact that you actually
39:17generate these like pretty detailed HTML
39:19reports and read them, I think it's
39:21probably the exception, dude. I think
39:22most people don't do this. Yeah. I mean,
39:25I think it's something I'm trying to
39:26push. Yeah. But like I think that like
39:28this is how you can learn to make just
39:31like I think the question for you is
39:33like okay like how do you go from making
39:35good shorts to making like really like
39:38the best shorts like highest production
39:40quality shorts out there. You know what
39:41I mean? And I think if that's the goal,
39:44right? Like I think you need to like
39:45then sort of probably learn more about
39:48like both about video production and
39:50editing and uh and like the technical
39:52concepts, right? And and then push
39:53yourself that way. And I think that like
39:55I think we all want to push ourselves
39:57towards like being better and faster and
40:00not just faster. Yeah. So,
40:01>> okay. So, maybe I'll add uh some custom
40:03instructions to my cloud.
40:05>> Be like, hey, make sure you generate
40:06HTML reports for everything so I can
40:08read that.
40:08>> Yeah. Yeah. Yeah. Well, I mean, you
40:10know, yeah, I think it's uh something
40:12you want to push yourself to as well,
40:14you know, to like figure out when are
40:16when do you want to understand how
40:17something is working.
40:18>> Got it. Awesome, man. Thank you for
40:19giving the gift of cloud code to all of
40:21us and um yeah really excited for what
40:24what's next dude. I don't know if you
40:25probably can't share anything but you
40:26know really excited for what's next. A
40:28lot of what's next is cloud tag you know
40:30and it'll keep getting better but I I
40:32think it's just something that it's hard
40:34for us to understate like how much it's
40:37sort of changed work at anthropic here
40:39and so excited for everyone else to to
40:41get on board or to try it out. Yeah,
40:42>> if you have a really capable employee
40:44you're not going to like micromanage
40:45them, right? You just put, hey, tag them
40:47on Slack and I can't do this. Then
40:49hopefully it's done.
40:50>> Or you want to like talk talk with
40:52iteratively with other people and
40:54collaborate. Yeah,
40:55>> you should build it to a point where I
40:56can walk by Claus's desk and just like
40:58ask some questions respond.
41:00>> Yeah, like a robot. Yeah, I mean that
41:02could be a hack project for you. You
41:03know, I feel like I could already do
41:04that. Yeah.
41:05>> Cool. All right, well um really great
41:07chatting and um Yeah, I I I think people
41:10know where to find you online, so I
41:11don't think we have to talk about that.
41:12>> Yeah, sounds good. Sounds good. Yeah.
41:15Amazing. Thanks, Peter. Appreciate

About this transcript

This page contains the full transcript of How I Plan, Build, and Run Loops with Claude Code in 40 Minutes | Thariq Shihipar by Peter Yang, generated from the public captions YouTube serves with the video. The transcript has 8,827 words across 1,218 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.