YT YouTube2Text

YouTube2Text → Transcripts → I Turned Claude Into a 24/7 Trader

I Turned Claude Into a 24/7 Trader — Transcript

by Nate Herk | AI Automation · 8,434 words · 1,137 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00So, Claude Opus 4.6 is finally here, and
0:02it got me thinking because as I was
0:03scrolling down and I started looking at
0:04the benchmarks, I noticed right here we
0:06have a genetic financial analysis, and
0:08we have about a 4% jump from Opus 4.6.
0:11And about a week ago, I dropped this
0:12video where me and Simon traded stocks
0:15for 30 days with our open Claude agents,
0:18and my open Claude agent was set up with
0:20Opus 4.6, and I was able to beat the S&P
0:22by about 8% as you can see right here.
0:24This green line is my chart, and this
0:26gray line is the S&P. So, if you guys
0:28want to check out this full video, I
0:29will link that right up here. So, after
0:30this challenge, I wanted to keep my
0:32agent running to see if it could keep
0:33beating the S&P, but now I want to
0:35upgrade it to 4.7 instead of 4.6. So,
0:38that's exactly what I want to cover with
0:39you guys today. How do we build a 24/7
0:41AI trading agent with Opus 4.6 inside of
0:44Claude code? We're not even going to
0:45touch open Claude or Hermes agent. We're
0:47going to just do this with Claude code.
0:49And this is all going to be thanks to
0:50another feature that we just got from
0:51the Anthropic team called routines. If
0:53you don't know what that is, I'm going
0:54to break it down in this video, but I
0:55also made a full one if you want to
0:56check that out, also right up here. So,
0:58basically, this is the unlock. 4.7 plus
1:01routines means you can have a 24/7 AI
1:03agent, and 4.7 was built for full
1:05throttle agentic work, judgment over
1:07ambiguity, and self-verifying outputs.
1:10So, it's perfect for these types of
1:11routines, and that's how I'm going to
1:12set up my 24/7 trading agent with Claude
1:15code routines. So, if you follow along
1:17with this video, you'll be able to
1:18pretty much duplicate my exact setup and
1:20have yourself a working agent and start
1:23trading with it. So, let's go ahead and
1:25get started. What are we building? We're
1:26going to create a Claude code project
1:28that runs on a schedule, something like
1:30a pre-market cron, a market open, a
1:32midday, and a close. It's going to
1:34research the market. It's going to place
1:35trades via the Alpaca API. It's going to
1:38journal and make sure that it's writing
1:40context into the file so it can learn.
1:42And then it's going to send me every
1:43day, every day of the week, an
1:44end-of-day summary to my ClickUp. The
1:46goal, obviously, is just to beat the
1:48S&P. We're not looking to do any crazy
1:49crypto or crazy day trading stuff. I'm
1:52just kind of doing this as a fun
1:53long-term investment. Just beat the S&P
1:55challenge. Like I said, in my first 30
1:57days I gave Opus $10,000 and we were
2:00able to beat the S&P by about 8%, which
2:02is really good, especially in such a
2:04short amount of time. So, what is the
2:06tech stack? We're going to be using
2:07Cloud Code Routines, that's kind of our
2:08scheduler. For the AI model, we're going
2:10to use Opus 4.7. For the skills, we're
2:12going to be using just kind of like
2:14building our own as we go. It'll be
2:15stuff like how the agent does research
2:17or how it makes decisions or how it
2:19places trades or, you know, makes logs,
2:21things like that. We're going to use the
2:22Alpaca API as our brokerage. We're going
2:25to use the Perplexity API for research.
2:27You could also just use the native web
2:28fetch and web search, but I'm going to
2:30use Perplexity here. And then I'm going
2:32to get my notifications in ClickUp this
2:34time rather than Telegram, just cuz I
2:36live in ClickUp a lot more. So,
2:37obviously some of these things can be
2:38interchangeable for you, but if you want
2:39to just copy my exact setup in this
2:41video, then feel free. Okay, so how to
2:43get started with these different tools
2:44that we need for this video? The first
2:46one is going to be Alpaca. So, go to
2:48alpaca.markets.
2:49You can see up at the top we even see
2:50want to trade using AI. We launched an
2:52Alpaca's MCP server to help you trade
2:54with natural language. So, you could use
2:55that. I'm just using the endpoints. I
2:57honestly like to use endpoints a lot
2:58more. But anyways, you're going to sign
3:00up right here and you'll click on
3:01trading API. Now, once you get in there,
3:04you will have basically a paper trading
3:05account and right here you can see I've
3:07got a paper trading account with 100k.
3:08You could also open up a new account
3:10right here and that's how you could put
3:11real money in. You might, like I said
3:13though, have to just verify your
3:14account, which could take a couple days.
3:16And then they're both the same as far as
3:18the credentials. You basically scroll
3:19down and on the right-hand side you have
3:21your API keys right here. You get an
3:22endpoint and you get a key, but you also
3:24need like a secret key. So, you have
3:25like a secret key and then you need like
3:26an ID. So, you'd hit regenerate,
3:28generate new keys, and this is where you
3:30get your key and your secret. And
3:33once you copy and paste this, just make
3:35sure to save it because then it's going
3:37to go away and you'll only be able to
3:38see the key, but you need both of them.
3:40Next up I have ClickUp. This is where I
3:41want to get my notifications. You don't
3:43have to use ClickUp, you can use Slack,
3:44you could use, you know, whatever,
3:46Telegram, WhatsApp. I would go up here
3:48and go to my settings, and then I scroll
3:50down and find ClickUp API, and then I
3:52would just copy this token right here.
3:54And then similarly with perplexity, I
3:55would go down here and I would click on
3:58all settings, I think. And then in here
4:00I can see API platform and this is where
4:02I would be able to go find my actual API
4:04key. If I went down to right here API
4:06keys, I can make a new one or copy an
4:08old one. But basically, the way it works
4:09is every platform has typically API
4:11keys. You just have to find them in the
4:13settings somewhere. If you can't find
4:14them, just do a quick perplexity or
4:15Google search or even ask Cloud Code and
4:17it should be able to help you get to
4:19your API keys. And now that you know how
4:20to get these keys, just hold on to them.
4:22We're going to actually plug them in
4:23later in the video. And then the final
4:25piece of the puzzle is going to be the
4:26Cloud desktop app. You can see here you
4:28can have Cloud chat, you can have Cloud
4:29co-work, or you can be in Cloud Code.
4:31And this is where we're going to set up
4:33these actual routines where now if I go
4:35to my calendar, you can see every single
4:37day I've got Bull, which is my trading
4:38agent, waking up at 6:00, 8:30, um noon,
4:433:00 p.m. And then on certain days like
4:45on Fridays, he's doing a weekly review.
4:46So this is the way I currently have mine
4:47set up and we're able to manage
4:49everything right from here inside of the
4:50Cloud desktop app because of this whole
4:53routine feature. So all you need to do
4:55to use the Cloud desktop app is just
4:56type in Cloud desktop app download and
4:58you should be able to install it right
4:59here. You go to the downloads page and
5:01then you just have to choose the Cloud
5:03desktop for your operating system. And
5:05then the only other thing you need to do
5:06is make sure that you are on an actual
5:08subscription. So it has to be a paid
5:10either 20 bucks a month plan or the max
5:12plans. Now in this video, you might also
5:14see me use Cloud Code inside of
5:15something that looks like this, which is
5:17the VS Code IDE and then we use Cloud
5:20Code extension inside of it. It's not
5:21mandatory, but that is where I typically
5:23prefer to use Cloud Code because I can
5:25see all of my files over here. So if you
5:27want to do that, just type in VS Code in
5:28your browser, open this up. It's
5:30completely free to download and then go
5:31ahead and download it and connect your
5:33Cloud Code account. So that is pretty
5:35much the text stack that we're going to
5:36be using today to get all of this set up
5:38and let's get back to the video. Okay,
5:40so before we get into the actual
5:41step-by-step kind of build, we are going
5:44to need to talk about the mental model.
5:46So understand this before you write
5:48anything. The trading strategy is a
5:50piece of the work, but the memory
5:51architecture that you're going to use is
5:53going to be huge.
5:54So, every time a routine fires at 7:00
5:56a.m., Claude code basically wakes up
5:58essentially stateless. It doesn't really
6:00know anything. So, how do you make a
6:01stateless agent act disciplined and
6:03remember rules and learn over time? You
6:05do that with files and with context. So,
6:08every routine basically wakes up, reads
6:11files,
6:12does the job, and then writes back any
6:15important lessons or anything in like
6:16that the next agent when it wakes up
6:18needs to know, it just writes that down.
6:20And that's kind of at a very high level
6:21the type of framework that we're going
6:23to be using here. So,
6:24the final thing I wanted to talk about
6:25before we get into this is the context
6:27budget. So, treat tokens like money,
6:30obviously. Every file that you ask the
6:31agent to read it's going to cost you.
6:33Each routine gets about 200,000 tokens
6:34to work with. I mean, we do have the a
6:36million token context window models now,
6:38but I don't really like to think of it
6:40like I have a million tokens to play
6:41with because context rot. So, system
6:44instructions will take some up, strategy
6:45files, your trade log, your API, your
6:47research. Every run is going to take a
6:49different amount of tokens, but just
6:50keep that in mind. And the reason here
6:52that we're using, if I go back to the
6:54stack, Claude code instead of just using
6:56something like, you know, the agent's
6:58SDK or I mean, I guess that could work,
6:59but rather than just using like a
7:01standard automation, is because if we do
7:03this with routines, we basically get the
7:06full autonomy of us working with Claude
7:09code manually because it goes through
7:11that whole agentic earnest loop of
7:13figuring things out. And especially with
7:14Opus for that 7, it's apparently going
7:16to be really, really good.
7:18Whereas, if you just had like an
7:19automation that would do research and
7:20then, you know, make a trade, you're not
7:22getting that super variability that you
7:24probably want with something that could
7:26be very volatile like trading. All
7:29right, so now that we got all of that
7:30stuff out of the way, let's get into the
7:31strategy. So, step one,
7:34think of this like you're teaching a kid
7:35to ride a bike. You have to explain how
7:38to balance before you go. You have to
7:40explain to the kid that he's going to
7:42steer with his hands and he's going to
7:43pedal with his feet. And even once you
7:45put him on the bike, you're not just
7:47going to let him go. You're going to
7:48hold the handlebars, you're going to put
7:49your hand on his back, and you're going
7:50to walk with him until he gets a little
7:52more comfortable. And then slowly you
7:53might take off, you know, you might stop
7:55steering, but you might just run behind
7:56him. And then slowly you'll put on some
7:58training wheels. And like you just keep
8:00bumping it up in phases. You can't
8:01expect to just throw a kid on a bike and
8:03it's instantly going to be magic. So
8:05maybe you want to start with paper
8:06trading. This is obviously not financial
8:08advice and I'm not, you know,
8:09recommending that you go put all of your
8:10life savings in here and expect to make
8:12money and then get mad at me. This is an
8:14experiment for me and I'm trying to show
8:16you guys what's possible. So just keep
8:17that in mind. Start with paper trading
8:19first if you're not comfortable. But
8:20anyways, the idea is you need a
8:22strategy. So for me, what I'm going to
8:24do is I'm basically going to extract all
8:26of the strategy that we've been using
8:27from my Open Claw agent, and I'm just
8:29going to say, "Hey, let's take this over
8:31to Cloud Code." So for me it's a little
8:32easier. But for you, if you actually are
8:34like a manual trader and every day you
8:36think about things and you do research
8:37and you place trades or you sell things,
8:39then just write down your strategy. How
8:41often are you checking the news? What
8:43are the signals that make you realize
8:44I'm going to buy this or I'm going to
8:45sell this? As much of your gut intuition
8:48and your routine about trading that you
8:50can get into this Cloud Code project,
8:52the better. And then I will over time
8:53you're going to help it learn from its
8:55mistakes, you're going to iterate, and
8:57it's going to get better and better
8:58every time you use it. But if you're
8:59starting completely fresh, which in my
9:02case of the Open Claw trading challenge,
9:04I literally just said, "Hey, I'm doing a
9:0730-day challenge. I'm trying to compete
9:08against my friend Salmon. Go to
9:10research. Spin up a team of sub agents.
9:11You're my wealth advisor. Try to beat
9:13the S&P. Figure out the best plan
9:15possible." And that's like genuinely all
9:17that I said to it, which I think is
9:18pretty cool. And it went off and it did
9:20a bunch of research. It figured out a
9:21strategy for me. And then for the next
9:2330 days it just did that. And I barely
9:24had to touch it at all.
9:26So if you want to start off by playing
9:27around with some research like that,
9:28then certainly do that. But if you guys
9:30remember at the beginning of this video,
9:31I kind of alluded to this benchmark
9:33here. So in the official release, and by
9:36the way, I just dropped a full video on
9:38this new release. I'll tag that up here.
9:39So, in the official release of Claude
9:41Opus 4.7, they always drop the
9:42benchmarks and they compare it to other
9:44models that are kind of leading in the
9:46industry as well. And so, this is the
9:48one I wanted to talk about in this case.
9:49We have a genetic financial analysis
9:5164.4%.
9:53We also have stuff like gradual level
9:55reasoning and you know, just the ability
9:56to think really well. Um genetic search
9:58actually is worse than Opus 4.6, which
10:00is interesting. But yeah, genetic
10:02financial analysis. So, what does that
10:03actually mean? This basically means that
10:05it is really good at analyzing a
10:07company. But it's not basically a signal
10:10or a benchmark for how good this model
10:13is at trading and specifically day
10:15trading. Like this doesn't mean that
10:16it's going to be able to look at like a
10:17technical analysis and look at the
10:18candlesticks and the MACD and figure out
10:21the exact best moment to buy and then
10:22the exact moment to sell. I'm not going
10:24to try to day trade with these agents. I
10:26just don't know enough about it to do
10:27it. But that's not what this metric is
10:29saying. This benchmark rewards models
10:31that can digest filings and write
10:32coherent fundamentals-driven theses or
10:35thesis, I guess. And that maps to
10:37long-term or swing or
10:39fundamentals-driven strategies, not day
10:41trading. So, that's why once again,
10:42we're just going to be trying to beat
10:43the S&P here. Okay, so that is the
10:45overall strategy. Let's talk about the
10:47scaffold. So, what you need to do to
10:49spin up the Claude code project in about
10:5160 seconds is you're going to open up a
10:53new folder and we're going to open up
10:54Claude code. So, let me hop over to my
10:56project right here. This is a new
10:58folder. It's called trading I think
11:00trading routine. This is basically just
11:02a settings file. There's hardly anything
11:03in there. The Claude doesn't D is blank
11:05and then we have a read me just because
11:07I turned this into a GitHub repo so I
11:08can pull it into a routine later. If you
11:10haven't done that yet, it's fine. We'll
11:12talk about that in a sec. The reason why
11:14we're starting here in VS Code is
11:15because I want to get started in a place
11:17where I can see all of my files on this
11:18left-hand side. Because if you're in the
11:20desktop app, which we will move over to
11:22here later, you don't get to see all
11:23your files open on the left-hand side.
11:25On the left-hand side, you see your
11:26routines and your other sessions, which
11:28is great, but when we're first getting
11:30set up, I like to be able to see
11:31everything in one spot. So, we're going
11:33to start here in the VS Code um either
11:35terminal or the Cloud Code extension.
11:38So, let me show you guys what I did.
11:39This is my Telegram chat with Bull, my
11:41trading bot, and I said, "Hey, I want to
11:43completely migrate you to a different AI
11:45agent setup. I need you to break down
11:46the strategy, the cron jobs, what are
11:48you looking for when you're doing your
11:49research, what type of sub agents do you
11:51have, what are the different signals,
11:53give me the full breakdown, all of your
11:55learnings." And so, it gave me all this
11:57stuff. And so, what I'm going to do is
11:59I'm just going to basically migrate over
12:00this strategy into my Cloud Code
12:02project. You can see we have all of
12:04these sell signals, we have all of these
12:06key learnings, and these kind of main
12:07frameworks, and we have our current
12:09portfolio. Real quick, guys, when it
12:11comes time for you to get set up, I have
12:13this free resource for you. It's a
12:1513-page PDF, and I am assuming if you
12:17dropped this thing into Cloud Code and
12:19then started to brainstorm with it, it
12:20would help you out a lot as far as the
12:22infrastructure and the folder structure
12:24and the way you want to get things set
12:25up and the way you want to create your
12:26routines and things like that. So, it
12:28should be very helpful. This you can
12:29download for completely free in my free
12:31school community. The link for that is
12:32down in the description. Once you get in
12:34there, just head over to the classroom,
12:35click on YouTube resources, and you'll
12:36be able to find it right in there. So,
12:39let's get back to the video. So, if I
12:40pull that up real quick, you can see
12:41here's today. We had a bit of a a jump,
12:43and then we came back down today. Here
12:45is the past month, so this is where we
12:47were with the um open cloud agent. But
12:49you guys are going to want to go to
12:50Alpaca and make an account, especially
12:52if you want to put real money in. You
12:53might have to take a few days for your
12:55account to get approved. So, head over
12:56there and get an account. Then I also
12:58told it to give me a technical cheat
13:00sheet as far as like the different API
13:01end points it uses and how it makes
13:02those requests, as you can see, just
13:04because it already has all this
13:05knowledge. So, why waste time trying to
13:07train a new agent if we could just
13:08migrate over all of these memories and
13:11knowledge to this new agent. And then at
13:13the end of that, it even gave me an
13:15export of all of these important files,
13:17like the agent instructions, the trading
13:18strategy, the trade log, the research
13:20log, the weekly review. And I'm just
13:22basically going to migrate all this
13:23over. So, let me show you what that's
13:24going to look like. So, you can see that
13:26it gave me this zip file of all of those
13:29seven files and I'm just basically going
13:30to extract those into this project right
13:32here. So I'm choosing the trading
13:34routine project, select folder and then
13:36extract and we should see all those
13:37files just popped up right over here on
13:39this left-hand side. Now what I'm going
13:40to do is just talk to my Claude. I'm
13:42going to switch over here to plan mode
13:43and I'm just going to start talking. All
13:45right, so I am creating you as a 24/7 AI
13:49agent trade bot for me. I had a project
13:51set up with open Claude and
13:54we were trading, you know, $10,000 using
13:56Alpaca for the past 30 days and we were
13:58doing pretty well, but now I want to
14:00move it over here so we can set you up
14:02as a Claude code routine. So I just
14:04dropped in a ton of context for you.
14:05There's a ton of files in this project.
14:07Um if you want to just take a second to
14:09get sort of like
14:11acclimated to what I dropped in there.
14:12There's going to be a lot of
14:13information, but basically
14:15that's like all of the knowledge and all
14:17of the strategies and all of the signals
14:18that our open Claude agent had and how
14:20it was able to be successful and I just
14:23want to migrate that over here to you so
14:24that we're not losing any of our
14:26knowledge. I'm also going to paste in
14:27some of the key things that it told me
14:29that you should make sure to know. So
14:31I'm going to drop in that stuff right
14:33below. Now some of this stuff you can
14:34take with a grain of salt because it
14:35might have been set up and optimized for
14:37an open Claude agent harness, but you're
14:40Claude code. You understand the best way
14:42to set up your projects and to do
14:44things. So what's really important is
14:45that you're ingesting this information
14:47and you're understanding it and then
14:49you're helping yourself organize this
14:50project in a way that makes the most
14:52sense to you. Okay, so I know that that
14:54is a ton of information that I just
14:56chucked into there and you guys might be
14:57sitting there like, okay, well I don't
14:58have all these files. What do I do? This
15:00is where you just start to brainstorm.
15:01Literally treat this thing as if it is
15:03your best friend who is the best trader
15:05in the world and you are just trying to
15:06give it your strategy now. So if we
15:08think back to earlier in step one where
15:09I talked about the strategy, when you
15:11open up a piece of paper or sorry, you
15:13take out a piece of paper, you open up a
15:14Google Doc and you just start writing
15:16down. How do you actually trade? What's
15:19your approach? Once you write down all
15:21this stuff and you basically just brain
15:22dump about how you want to do it, then
15:24you plug it into the Claude code, and
15:25that's where you see how I set up this
15:27project here. I'm just helping it get
15:29situated with what we want to do right
15:30now. And honestly, the more context and
15:32details that you give it right now, the
15:34more it asks you questions, the better.
15:36As you can see, I'm using the
15:37brainstorming skill from Superpowers,
15:39and it's going to start off by asking me
15:40some questions. Okay, so I just answered
15:42some questions, and I did want to show
15:43you guys this. So, it alerted me about
15:45something with security inside of this
15:47template that Open Claude gave me, it
15:50had my live Alpaca key in there. So, I'm
15:53going to go ahead and rotate those
15:54before we actually push this new one
15:56into production, but just something to
15:58be thinking about when you're migrating
15:59stuff or when you're feeding in API keys
16:02directly to a chat. Just be careful of
16:03that kind of stuff. And I'm going to
16:04show you how we get our API keys set up
16:06in routines because you don't want to
16:07put them straight into a GitHub repo.
16:09You're going to have to set them up in
16:10the environment variables, so I will
16:11show you guys that as well. Okay, so
16:13this point it's come back with a plan.
16:15So, Nate ran a live money Alpaca trading
16:17bot called Bull on the Open Claude
16:19harness for 30 days, and now we have all
16:21these context files with strategy, with
16:23trade history, with research notes, blah
16:25blah blah. The goal of this migration is
16:26to make Claude code a first-class home
16:28for this bot. So, we will set up the
16:29cron scheduling, the notifications.
16:31We're getting rid of that agent mail
16:32step, so me and Solomon were having our
16:34agents trash talk each other. We're
16:36going to rotate our credentials, of
16:37course. And then it's proposing a layout
16:39for the way that the files should be set
16:41up. So, you can see, it's going to put
16:42the files in here into the memory, and
16:45then we're going to be able to prompt
16:46our routines to always check the memory
16:48first and update the memory. So, this is
16:50where it will really help for you guys
16:51to brainstorm with how the project
16:52should be set up. Oh, one other thing
16:54that I forgot to give it here is that I
16:55want it to use Perplexity for research.
16:57So, I'm going to come down here to the
16:58bottom of the plan, and I'm going to
17:00say,
17:01"Yeah, this all looks pretty much good
17:03to go, but one thing I forgot to tell
17:04you is that we want to use Perplexity
17:06for research rather than just the native
17:08Claude code web search and web fetch
17:10function. So, add in that Perplexity
17:12will be used for web search, and once
17:13again, once we set up up routine, that's
17:15when I'll give you the Perplexity API
17:17key. Okay, so after this change has been
17:18made, we should be ready to start
17:20migrating over to setting up the
17:22routines. All right, so the new plan has
17:24come back. So, here's the end-to-end
17:26check. Secrets, Alpaca, memory file
17:29commands, dry run, ClickUp, Perplexity,
17:32nice. We're going to go ahead and accept
17:34these changes. And you also notice that
17:36we have auto mode on. So, I think with
17:38this new update with Opus 4.7, when you
17:40auto-accept a plan, it turns on auto
17:42mode, which means it lets Cloud handle
17:43permission prompts automatically. Now,
17:46you will have to be careful though
17:47because doing this is a little bit more
17:48expensive. So, if you don't want it to
17:49be auto mode, then you can just switch
17:51it right off of auto mode, of course.
17:53But, um yeah, just something I wanted to
17:54point out. All right, so while all of
17:56that's still getting set up, let's talk
17:57about step three, which is guardrails.
18:00So, before you start working in the
18:01trading logic, think about this kind of
18:03stuff. I obviously talked about how you
18:04could start with paper mode, and then
18:06you toggle on real trading and put in
18:08money when you feel comfortable. You
18:10could do things like, "Hey, max 5% of my
18:12portfolio per position." You can have a
18:14daily loss cap. You can tell the agent
18:16not to do, like only buy three new
18:17positions per week, or no options ever,
18:19or don't do this, or always do this.
18:21Because the agent is going to be
18:23autonomous, it's going to be eager, and
18:24if you don't give it guardrails, it
18:25might just kind of start to, you know,
18:27go off the rails. But like I said, even
18:29though I'm migrating a system over, I
18:30know it's not going to be perfect. So,
18:32I'm going to watch every single run, and
18:33I'm going to read through the
18:35conversation history of every run, and
18:36I'm going to see if there's things that
18:37I need to tweak in the prompt, or things
18:39that I need to tweak in the project
18:41settings, or the context files. And like
18:43I said, just kind of keep building the
18:44plan as we're flying it here. And then,
18:46that's where we also start to develop
18:48skills. Because right now we have skills
18:50maybe for like a research skill, or for
18:52a trade skill. But we're going to make
18:53sure that we're getting them actually
18:55explicitly invoked, and we're going to
18:56be making sure that that increases the
18:58consistency, and we're going to make
19:00sure that the agent is able to give us
19:02insights every day of, "Hey, here is
19:04what worked well, and here's what I
19:05noticed, and hey, here's where I think
19:06that maybe we should update these skills
19:08X, Y, and Z, you you whatever." So, a
19:11combination of thinking about the
19:12guardrails and thinking about skills is
19:14how you can continuously make this thing
19:16better over time. And now that we've
19:17talked about all that, we will get to
19:18the routines and the deployment. Let's
19:20check in on our
19:22guy right here. And because I turned it
19:23off auto mode, it was sitting and
19:25waiting for me to do a permission. All
19:27right, so everything has been migrated.
19:29We're looking good, but we're getting a
19:31little bit to the point where I like to
19:32reset the context. So, before we move
19:34on, I'm going to say, "Hey, give me a
19:36summary of what we've done." And then
19:37I'm going to clear the context and we'll
19:38get started again. All right, so here's
19:40the session summary that it gave me.
19:42Copy that.
19:44Do a slash clear.
19:46Paste that in, hit enter, and now we're
19:47basically right back at the beginning
19:49with a fresh window of context. So, this
19:52is where it gets pretty cool. You can
19:53see on the left-hand side it has
19:54reorganized things. We have our scripts,
19:56we have our memory files, we even have
19:58some commands in here, and it built all
19:59of that just because I told it that I
20:01was trying to migrate stuff over. It
20:02also updated this Claude.md file because
20:05now we have Bull back in all of his
20:08glory, and we have our Claude.md file
20:09that is about 156 lines. So, now it's
20:12time to start planning out step five,
20:14which is the routines. So, we basically
20:16need to figure out what are the
20:17different daily triggers that we
20:18actually want. And we also don't maybe
20:20want them every day because the market's
20:21only open five days a week. So, this is
20:23kind of like a baseline. This is kind of
20:24how it worked with my Bull trading
20:26agent. But I'm actually just going to
20:28come into VS Code and we're just going
20:30to brainstorm once again. So, I'm going
20:31to go on plan mode and I'm just going to
20:33start asking. All right, so now that you
20:34know everything about this project and
20:36you understand what we did with Open
20:38Claude, I want you to help me figure out
20:40what is the best way that we should be
20:41setting up these Claude code routines to
20:44basically simulate that exact same type
20:46of environment. So, we have those cron
20:48jobs, but each one is kind of doing a
20:50bit of a different thing. You know, the
20:51pre-market one does a bit of a different
20:52job than the mid-market one. And the the
20:55day close one does a bit of a different
20:57job. So, you have a better idea of all
20:59of our memory files and our skills, and
21:02you understand the structure of every
21:04time the agent wakes up and the routine
21:06is triggered, they need to read the
21:08files first to, you know, kind of get it
21:09acclimated with the environment, do what
21:11they need to do, research, potentially
21:13make a trade, potentially sell
21:15something, you know, whatever it is
21:16based on their best judgment, and then
21:18update all of the memory files so that
21:20next time the next agent wakes up, they
21:22understand what's going on. So, what I
21:23want you to do is help me figure out
21:25what are the right structure for these
21:28um
21:29cron jobs, and
21:31most likely we only want them during the
21:32week when the market's actually open.
21:34And just like that, I'm able to brain
21:36dump and Claude's going to do the heavy
21:37lifting for me and propose some ideas.
21:40Okay, let's see the plan that it came up
21:41with. So, stand up five recurring Claude
21:44code schedule triggers, and we've got
21:46these different slash commands that it's
21:48going to be able to utilize. So, all
21:50five triggers will run on the weekdays
21:51only except for the weekly review, which
21:53is Friday only. So, we have a
21:54pre-market, which is for research
21:56catalysts and draft trade ideas. It
21:58won't message me unless it's urgent. We
22:00have a market open, which is going to
22:01execute the planned trades and set 10%
22:04trailing stops that'll go to the trade
22:06log. Only if a trade is placed will I
22:08get a notification. We have midday,
22:10which will cut minus 7% losers and
22:12tighten stops on winners, also going to
22:14trade log. So, here you guys can see
22:15basically it was able to analyze
22:17our strategy and analyze what we have
22:19access to and then build me a plan. And
22:21now what I want to do is basically just
22:22have it create those prompts for me so
22:24that I can go ahead and go set up those
22:26routines in Claude code. All right, so
22:28those crons look great. So, I don't need
22:31you to actually create them. What I need
22:33you to create for me is the prompts. As
22:35if I was going to run one of those
22:37commands into a fresh Claude code
22:38session and I wanted it to do all the
22:40right stuff, what is the exact prompt
22:42that I would enter into that session?
22:44And the one thing that you need to keep
22:45in mind here is reminding it to
22:47obviously um read the files before it
22:49does anything, and update the files at
22:51the end, and then also when it needs to
22:53use an API like Perplexity or ClickUp or
22:55Alpaca, that it should be grabbing those
22:58keys from the environment variables of
23:00that environment, not inside a .env
23:03file. All right, and now with that
23:05feedback, it should be able to give us
23:06those prompts, and then we should be
23:07able to just set up the routines and go
23:09ahead and test them out. Okay, so while
23:11this is writing up, there is one more
23:12thing that I wanted to show you guys
23:13that is very important. So, in the Cloud
23:15Code desktop app, when you go to
23:17routines and you go to create a new one,
23:18you can see that you can create them
23:20locally or remote. So, if you create
23:22them locally like some of these I have,
23:24this just basically means that it lives
23:26in your local directory. It can still
23:27talk to the internet and do all those
23:28things, but it lives in your machine
23:31essentially. And that means if you
23:33closed out of this desktop app, that
23:35local
23:36routine would not fire off. And that's
23:38why these remote routines are so cool
23:40because your computer can be off, but
23:41they're still running because they run
23:43in the cloud. But what that means is
23:45when you do a remote run, you have to
23:47have a GitHub repository to have this
23:50run in, which is why I have this trading
23:51routine repo set up.
23:53And so, when they run via a cloud, you
23:56know, routine, it basically clones that
23:59repo in the cloud. It spins up a copy.
24:01It works out of it. It does what it
24:02needs to do, and then it basically like
24:04destroys that cloud environment. So, if
24:05you want them to run remotely and you
24:07want those files to persist every
24:09session, you just have to talk to Cloud
24:10Code and say, "Hey, make sure, you know,
24:12because we're doing a routine, make sure
24:14that all of these files that it's
24:15actually updating, it's able to push
24:17back and commit back to that main actual
24:21branch, back to that main." And we just
24:22got a notification from my trading bot
24:24because it was market close. But
24:27anyways, I was saying, make sure that
24:29it's able to push those changes back to
24:31the main repo. Otherwise, the next
24:32agent's not going to pick it up, and
24:33then what's even the point? So, you can
24:35start local if you want, and then you
24:36can
24:37sort of graduate to the cloud if you
24:39want, but I definitely just wanted to
24:40make sure you guys understood that
24:42difference there. Okay, so that's done.
24:44You can see it created this new folder
24:45right here called routines, and this is
24:47basically the prompt that we'd be
24:48feeding into each of the different
24:50routines that we make in order to make
24:52sure that Bull actually understands what
24:54to do. And the last thing we need to do
24:56is actually push all of these new
24:57changes to this GitHub repo because
25:00right now here is my trading routine
25:01private repo. All that's in there is a
25:03readme and I don't think there's
25:04anything in here. So, we need to get all
25:05of these files that we just created, you
25:07know, these commands, these scripts,
25:09these memory files. We can just have to
25:10get all of this into that repo so that
25:12our routine can actually access it. So,
25:13all I have to do is say, "Hey,
25:15push all of these changes, commit
25:16everything to the repo." By the way, if
25:18you've never done that before, it's
25:19super super simple. You go to
25:21github.com, create an account. It's free
25:22to create an account. And GitHub is
25:24really just like
25:25um it's like working on um a Microsoft
25:28Word on your computer and that's like
25:30local. And if you wanted to share it
25:32with someone, you would have to put it
25:33into like the shared drive or, you know,
25:35a a Google Drive or something. It just
25:37means that you could grab a code base
25:39from a different machine because you've
25:41uploaded it somewhere on the cloud. So,
25:43that's kind of why we're doing this
25:44whole thing with the routines in the
25:45cloud. So, when you get to GitHub, you
25:47come here, you create a new repository
25:48or
25:49the easiest way to do it is you just
25:51come to your Cloud Code and you say,
25:52"Hey, help me turn this into a GitHub
25:54repo." And it will say, "Okay, I need to
25:56authenticate you. Can I use the CLI?"
25:58And you say, "Yes, go ahead and do
25:59that." And then it will basically open
26:01up a tab and it will say, "Hey, can you
26:03sign in with your GitHub?" You sign in
26:05and now your Cloud Code just has your
26:06GitHub credentials. And whenever you
26:08need to push changes or clone a repo or
26:11whatever you need to do with GitHub, it
26:12can just take care of it because
26:14GitHub's been around for a long time, so
26:16it knows it way better than you probably
26:17do. It knows it way better than I do.
26:19So,
26:20when in doubt, just ask Cloud Code. All
26:22right, so if I give this puppy a
26:23refresh, we should see that we now have
26:25all of these files popped in here. We've
26:27got bull trading routine. Awesome. So,
26:29let's head over to the Cloud Desktop app
26:31now and start getting set up. So, what
26:33you're going to do is you're going to
26:34click on new session and you want to
26:35open up the local projects right here of
26:38whichever thing you just set up there.
26:40So, right here you can see that I'm in
26:41the folder called trading routine, which
26:43we were just working in in VS Code. And
26:45if you guys remember, in VS Code, we had
26:47these five different routine prompts.
26:49And inside of each of these, they had
26:50the cron, they had the actual like
26:53steps, and then they had like any of the
26:55skills or commands that they need to
26:56run. And this is in a folder once again
26:58called routines. So, before we start
27:00talking to our new project, we need to
27:02set up an actual environment for this.
27:04So, if I go to new routine and I go to
27:05remote and I go over here to my um cloud
27:09environments, you can see that you can
27:10add one. And so, just add one and I'm
27:13going to pull up my trading one so you
27:14guys can see. But here, this is where we
27:16set up things like our API keys and our
27:17network access. So, I called mine
27:19trading. I gave it full network and this
27:22is where I'm going to put in my API keys
27:23for Alpaca and for Perplexity and for
27:25ClickUp.
27:27And then you basically just go ahead and
27:28save the changes of that environment and
27:30this is the one that you want your
27:32actual routines to be scheduled inside
27:35of. I am going to basically just talk to
27:37it right here to help us get these
27:39remote routines set up. Look inside the
27:42routines folder inside of this project
27:45and help me set up all of these
27:46different scheduled routine runs. They
27:49should all have a cron job in there.
27:50They should all have a prompt in there
27:52already and we will be working out of
27:54this GitHub repo and scheduling off
27:56these routines. So, let me know if you
27:57have any questions. Do not exit the
27:59planning phase until you're 95%
28:01confident that you know exactly what to
28:03do and exactly what I'm looking for.
28:05So, now I'll shoot that off and it
28:06should be able to come back with a plan
28:07to actually go create all of those cron
28:09jobs for us. Okay, now you can see it's
28:11coming back and asking us some
28:12questions. So, the first one is which
28:13cloud environment should they run in?
28:15We're going to do trading. That's the
28:16one you guys saw me set up. We want to
28:17use Cloud Opus 4.7 and we're going to
28:20use this GitHub repo. Yes, trading
28:22routine. All right, so it's set up all
28:24of those routines for us. You see that
28:26all I had to do was talk to it. It set
28:27up these crons cuz I don't know what
28:29exactly this means, but basically that
28:30means 6:00 a.m. Monday through Friday
28:32for all of these. Monday through Friday
28:348:30 a.m., noon, 3:00 p.m. and then this
28:36one is just 4:00 p.m. on Friday. So, if
28:38I go over here to my routines and I
28:40scroll down, you can see that we have
28:41these new five routines for bull and
28:43these are all remote. If I click in this
28:45one for example says every Friday at
28:474:00 p.m. Central, we're in the right
28:49repo. If I click into here, we should
28:51see we see this entire prompt, which is
28:53correct, and we're in the right
28:55environment, which is my trading cloud
28:56environment. And then one other
28:58important thing that you have to do is
28:59inside of your actual um routine, what
29:02you want to do is go to permissions, and
29:03you want to turn on allow unrestricted
29:06branch pushes. This removes the branch
29:08push restrictions, and Claude will be
29:10able to actually push to any branch, not
29:11just the Claude {slash} branches. So,
29:14turn that on for all five of these
29:15routines here. So, I'm going to turn
29:16that on for all five of these routines
29:18here. So, let's just pretend that it is
29:20Friday. I'm going to do a run now to see
29:22if this is able to actually work. And
29:23when you run these, you can actually
29:25click right here, and you can watch them
29:26actually work in real time. And best
29:28practice with routines is that every
29:30single time you build a new one, you
29:31should always do a run now at least a
29:33few times to make sure that they are
29:34working as expected. And exactly why we
29:36test out, you can see what happened is
29:38when it tried to hit Alpaca, it wasn't
29:40able to find the API key because for
29:41some reason, once again, it thought it
29:43was going to be in a {dot} env, but as
29:44you guys know, we don't have a {dot} env
29:46because we don't put that in the GitHub.
29:47So, I'm coming back into this session
29:49that set up these five, and I'm just
29:50going to say, "Hey, remember, all of the
29:52API keys are going to be stored in the
29:53environment variables. So, all of the
29:54prompts must explicitly state that."
29:56Okay, so the issue was they weren't
29:58spelled exactly word for word, letter
30:00for letter. So, I fixed that, and you
30:02can even see, for example, this one,
30:04they're all now coming through, but for
30:05example, in this one, I have the word
30:07API right here between Alpaca and
30:09secret. So, they have to be matching one
30:12for one, otherwise it will think that
30:13they're not there. But anyways, now that
30:15that has been figured out, it's actually
30:17able to look through all of these
30:18things, it's able to look through the
30:19memory, and it's going to execute our
30:21weekly review here. And there we go,
30:22just like that, the weekly review has
30:24been ran. I also told it that I fixed
30:25the Alpaca secret key, so now all of
30:27that's been updated. And if you go over
30:29to my GitHub, you can see that it was
30:31able to push these actual commits to
30:33this repo. This one was just a test one
30:35that it was running on the background as
30:37well, but it was able to actually give
30:39us the weekly review, and you can see it
30:40was committed by Claude, and it made
30:42changes to this actual MD file. So now,
30:45the next time it actually will be able
30:46to pick it up and we have our stuff all
30:48synced up. And alongside that weekly
30:49review, you can see that in my internal
30:51automations channel, sandwiched between
30:53two YouTube notifications, we have the
30:56weekly review that the run just sent us.
30:58We have our portfolio, we have versus
31:00the S&P, and we have any trades that we
31:02were making, and then, you know, some
31:03best and worst, and also it graded
31:06itself a C for the week. So, we'll see
31:08how this is able to improve. So, I just
31:10spun up this test so you guys could see
31:11that we are able to talk to Alpaca right
31:13now. You can see that we're able to get
31:15the current balance. And if I go into
31:17here, we should see 9859. And back in
31:20Claude Desktop, where it pulled it,
31:229859. And just like that, we have set up
31:24a 24/7 AI agent trade system with Opus
31:274.7. We've got our 6:00 a.m., our 8:30
31:29a.m., we've got our noon, and our 3:00
31:31p.m. And then on Fridays, we have our
31:33actual weekly review as well. Now, of
31:35course, like I talked about earlier, it
31:37is on you to now keep testing, keep
31:39iterating, making sure that everything's
31:40working, and upgrading your strategy as
31:42you go. And maybe you want to add some
31:43more crons in. Maybe you want to take
31:44some away. Maybe you want to add one on
31:46just like a Sunday night to do some more
31:48research to prep for the week. However,
31:49you would actually want to set up your
31:51trading strategy if you were doing it
31:53manually. Now, I know you guys are
31:54thinking like I have four automations
31:56per day basically plugged in now with
31:58Claude routines, and that's going to eat
32:00away at my session limits, right? I
32:02mean, yes and no. I'm definitely going
32:03to be watching it, obviously. But keep
32:05in mind, when I did this with my open
32:06Claude agent for 30 days, I was also
32:08plugged in my Claude subscription, and I
32:10was still doing my own workflows, and I
32:12was never really hitting my limits. It's
32:14all about managing your context. And I'm
32:16going to have a full video dropping
32:17about that, another one coming out soon.
32:19But I can tell you for certain that if I
32:21were to set up all of these crons and
32:22have the same sort of autonomy and
32:24agentic workflow-ness, and I was just
32:27doing this via API, then that would 100%
32:29be way more expensive than what I've got
32:32set up here. We have now deployed, and
32:33now we just go ahead and track. And what
32:35is the hidden lesson here? The hidden
32:36lesson is that files aren't just memory,
32:38but they're essentially the agent's full
32:39personality and discipline. Like with
32:41Open Claw, you set up your soul.md file
32:43and you've got your agents.md file and
32:45you got all these different files. So,
32:46it's really about how do you make sure
32:48that you orchestrate this in the right
32:49way so that it can access the right ones
32:51when it needs to. So, anyways, I hope by
32:53the end of this video you guys are able
32:54to follow along and set this up. I'm
32:56going to be giving away this document
32:57and it's going to have way more
32:58information in there basically about
32:59everything that we've gone over today
33:01and I'll drop that in my free school
33:02community. The link for that is down in
33:04the description. Once again, completely
33:06free. But, that is going to do it for
33:07today. So, if you enjoyed the video or
33:08you learned something new, please give
33:09it a like. It helps me out a ton. And as
33:11always, I appreciate you guys making it
33:12to the end and I'll see you in the next
33:13one.
33:14Thanks, everyone.

About this transcript

This page contains the full transcript of I Turned Claude Into a 24/7 Trader by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 8,434 words across 1,137 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.