§ 01 · The Hook
The bait, then the rug-pull.
Theo handed himself a challenge: hit the 2x holiday rate limit ceiling on Claude Code. What followed was two weeks of six parallel instances, zero IDE opens, 11,900 lines of generated code, and a $1,500 inference bill he paid $200 for — and it fundamentally changed how he thinks about writing software.
What the video promised.
stated at 01:25 "I am a little Claude Code pilled, and I wanna share a bit about why." delivered at 32:20
Where the time goes.
00:00 – 02:30
01 · Cold open & hook
Holiday rate limit 2x → decided to go all in. Six parallel instances, no IDE. Two full projects, OS config.
02:30 – 03:00
02 · Sponsor — G2i
G2i hiring platform sponsorship. First PR in 7 days guarantee.
03:00 – 09:00
03 · Image Studio demo
Prototyping playground for image gen UX. Sidebar scroll fix live. Multi-model gen (Nano Banana Pro, SeaDream, GPT Image). Image-to-chat follow-up feature. Built with Convex + Vite, all UI by Claude.
09:00 – 16:00
04 · Monorepo conversion
Sent 'convert to TurboRepo monorepo with React Native mobile app' as a joke. Plan took 20 min. Ran 1+ hour. Fixed its own errors. 2,300 lines added. Greptile 5/5.
16:00 – 18:00
05 · Auth across all platforms
Clerk auth added to web + mobile + Convex simultaneously. 1,800 lines. Another 5/5 from Greptile.
18:00 – 22:00
06 · Claude Blocker extension
Chrome extension locking Twitter unless Claude Code is running. Built in 30 min across 3 parallel instances. Open source.
22:00 – 24:00
07 · Using your computer differently
JJ config, zsh scripts, system setup via Claude Code. Not doing things he could not, doing things he did not bother with before.
24:00 – 26:30
08 · Claude Code Safety Net plugin
Plugin blocking destructive git/fs commands even in --dangerously mode.
26:30 – 28:00
09 · OpenCode comparison
Best multi-model option acknowledged. Did not use this run due to subscription workaround friction.
28:00 – 30:00
10 · Ralph Wiggum Loop
Bash loop feeding continue past done checkpoints. Mostly unnecessary with Opus.
30:00 – 32:20
11 · Complaints + usage math
Skills = markdown files. Hooks half-finished. At $200/mo hit 7% limit all out. ~$1,500 inference consumed.
Visual structure at a glance.
Named ideas worth stealing.
09:40
concept
Monorepo with TurboRepo + Convex
TurboRepo shares Convex backend between web and React Native. Convex is just a folder so models edit it directly without MCP dashboard access.
Steal for Multi-platform apps — eliminates dashboard-dependent MCP friction for backend config
07:10
concept
Design route comparison
Spin up 3 routes (/1 /2 /3) with the same feature, compare visually, cherry-pick best elements.
Steal for Fast UI iteration without committing to a design direction
12:30
concept
Front end design skill
- Markdown file telling the model: do not make it look like AI slop
A CLAUDE.md skill file that elevates UI quality without per-prompt instructions.
Steal for Non-generic UI from Claude on every project
28:20
concept
Ralph Wiggum Loop
A bash loop feeding 'continue' until Claude actually finishes, bypassing intermediate confirmation checkpoints.
Steal for Long-running autonomous tasks that pause between phases
24:50
tool
Claude Code Safety Net
Plugin (kenryu42/claude-code-safety-net) intercepting destructive git/fs commands even in --allow-dangerously mode.
Steal for Running dangerously mode with reduced data loss risk
Lines you could clip.
18:00
"It's changing whether or not I'm willing to make a project, not whether or not I'm capable. I was always capable of building things like this. I'm not doing things I couldn't do before. I'm doing things that I didn't bother doing before."
18:24
"These tools are significantly better if you already know how to code. Their strengths are much weaker if you don't yet."
18:13
"This is a fundamental change in how I think about writing code."
31:33
"I paid $200 and I got 1,500 out. As a user, I love this. As a company also selling inference that pays API pricing... you're welcome, you ungrateful bastards."
22:18
"I'm asking it to use my computer and make changes to my setup in my environment the way I normally would between like five different tabs, a bunch of searching, a bunch of trial and error, or I can just tell Claude Code to do it and go grab a tea."
How they spent the runtime.
Hook length85s
Info densityhigh
Filler8%
Sponsor blocks
- 02:30 – 03:00 · G2i
Things they pointed at.
How they asked for the click.
07:00
subscribe
"Hit that subscribe button, less than half of y'all are subscribed, we're so close to hitting 500k."
Mid-video subscribe ask, casual and brief
Word for word.
HOOK opening / re-engagementCTA the pitch metaphor analogy story
00:00HOOKOver the holiday break, something interesting happened. Anthropic gave you a two x increase in your rate limits on Claude Code. And I decided that I wanted to try my best to hit it. And man, have I had a wild ride. I never thought I'd see the day where I'm running six Claude Code instances in parallel, but this is basically my life now. I haven't opened an IDE in days.
00:21HOOKI have been building more than I've ever built. And I'm questioning what the future looks like for us as an industry. I knew Claude code had improved since I last used it and I knew Opus 4.5 was capable of things that I didn't think were possible using LLMs previously. I did not think this would fundamentally change the way that I write code. Over the holiday break, I built two full projects from scratch, a web and mobile app for one of them, did some crazy overhauls to those projects throughout, published a bunch of work for new features in t three chat and just
00:52HOOKconfigured my entire operating system using Cloud Code. I think it's safe to say at this point in time, I am a little Cloud Code pilled, and I wanna share a bit about why. That said, Anthropic is certainly not paying me any money. In fact, I had to bump the $200 a month subscription to do this, so we're gonna do a quick break for somebody who is paying me, today's sponsor. A lot of people say my products are only successful because I'm popular. And they're kind of right, but not for the reason you think. Turns out if you have a bad product and you put it in front of a bunch of people, they don't care. If you have a good product and put it in front of people, they do. But that's the key. You have to have a good product. In order to do that, you need a good team. This is the benefit I have from being popular. I have access to millions of the best engineers in the world. And I don't know how I would build without that, because I remember the days where I was recruiting the old school way, and it was hellish.
01:41If I didn't have access to the engineers I do, I would certainly be using today's sponsor, G2i. These guys know how to hire. In fact, they have placed some of the best people in my community at really cool companies. I've been blown away at the projects that G2I's involving themselves in. Everything from Webflow to 1Password to Facebook themselves. Yes, Meta hires through G2I now. And it makes sense why. They really understand what developers
02:05need to be successful. They run things like React Miami, the people who work there watch every video that we post, they know how to integrate with your team in a way that doesn't feel like some weird third party arm doing recruiting for you. They're building a platform to help you find the best people for the gig when you need them. It doesn't matter what you're trying to fill, junior, senior, front end, back end, full stack, mobile, data scientist,
02:27all the above, for all sizes of companies, by the way. If you just got out of your YC batch and you're trying to get a good engineer, don't waste all of your time going through recruitment hell when you can work with a partner like G2i and have a much, much better time. Their goal is to get the first PR in within seven days of joining, and I've seen this happen enough times to believe it. Stop wasting your time hiring, get back to building at soydev.link/t2i.
02:49I'll admit, this was a demo I just rigged up for the sake of this video. I just had a bunch of Cloud Code instances editing pages on my personal site. But it's not too too far from how I've been using things in reality. I have been making a lot of real work
03:06and doing a lot of surprisingly difficult tasks with Claude Code. I've gone so deep that even Ben is surprised. I'm doing things he didn't think the models were capable of. Just to get us started, I wanna show you guys the main project I cooked throughout break. This is my new Image Studio app. This is meant to be a prototyping playground to figure out what we want the image generation experience to be like in t three chat. And yes, it's that ImageGen
03:30mock demo I've been doing all the time. The difference is it's no longer a mock. So apparently, I never set up scrolling for the sidebar which is kind of hilarious. Normally, my screen's a little higher resolution than it is when I'm filming. So I can't actually hit the generation button because I never set up a scroll view here. So I'm just gonna show you how I've been using things. The sidebar in the web app does not allow me to scroll. This means that on small screens, I'm unable to hit the generate button. And now, we wait not even that long,
04:00and this issue will be fixed. You might have noticed I specified web app. There's a reason for that. Not only did I build this app as a fully functioning web interface, I actually went as far as to make it a mono repo with a mobile app too. And how I did that is possibly even crazier than the fact that I did both.
04:20We'll get there, don't worry. One of the things of note is that when you start a new session, it doesn't really keep track of any context. You can slash resume to go to an old session ID, but if you don't, it has to recontextualize its knowledge of the code base. It's actually been really nice to be in a long running thread where I just tell it to do something else and it knows enough from previously to know where to start. I no longer find myself clearing the history and starting a new chat thread for every single thing I do when I'm doing agent decoding with Cloud Code. Especially if it has a pretty good idea of where things are in a code base for a medium sized project like this. To be fair, the implementations on any of the pieces of this project are still relatively small, but the amount of service area there is to cover now across the multiple packages is actually quite reasonable, especially considering the fact that I have yet to open an IDE for this project.
05:11Forgot to turn on bypass permissions because I'm running it aloud dangerously now. That's the point I'm at in life. Unrun dev. Oh, it does not appear to have worked. When I try and scroll on the sidebar, it attempts to do a whole page scroll, which doesn't work because the only scroll container currently functioning is the scroll container for the grid.
05:32I think we need a different approach to make scrolling work properly on the sidebar. Of course, the thing isn't working the way it does almost always when I am using it, when I am now here trying to film. Chat making a very good point that this alone just disqualified new vibe coders. Yeah.
05:49This is still the problem. And I wanna make sure that when I do content about this stuff, that I emphasize this point. These tools are significantly better if you already know how to code. Their strengths are much weaker if you don't yet. But now that I've made this one specification, it was able to make the right changes and get this all working as expected.
06:06Awesome. Sidebar form fields now scroll independently when the generate button stays fixed at the bottom, and I will command minus a bunch to make sure it goes away when it's zoomed out enough. It does. Awesome. So now that we know this all is working as expected, I sure, we'll do 16 by nine. Why not? Standard resolution because some of these models don't support high res. Actually, if I turn off Nano Banana, we can do high res. Generate.
06:27CTAAnd now we have three models separately generating images for this random test I wanted to do. Of course, the back end is all through Convex. You guys know, I've been loving Convex. Front end is a standard v tail end app. All of the UI was done by Claude. We had a lot of back and forth. This is rendition like five or six, like full UI overhauls Because those types of things are surprisingly easy to do with a tool like ClaudeCode.
06:51CTAIf you just spin up a few work trees, throw Claude in them, and tell it, you can even copy paste the exact same prompt, like, want to redesign this homepage. Something stupid I have been doing recently is having it spin up three different routes with the exact same feature, like slash one, slash two, slash three, then I go and compare them and tell them, I like this in route one, but there are some things I like about route two and three. Can we bring these ideas over but focus on the design from route one? And it handles this stuff incredibly well. And good point, Chad, I did also use Chad CN. Chad CN has made my life much easier. By the way, if you like these types of exploratory videos about how the dev world is changing and how my own development work is changing, love doing this type of content, and one of the best signals for the team that you like this is to hit that subscribe button. Less than half of y'all are subscribed, we're so close to hitting 500 k. So if you don't mind, hit that button, helps us out a lot. But let's pick one of these that we like. This looks good enough. Obviously, arrow navigation works plenty, but I didn't even need to tell it to do that. It just did.
07:48Sea Dream did not honor the aspect ratio I gave it. That's a good thing to know. Banana Banana Pro did. Let's do one of the features I was most excited about here. You can't see this button, but there's a little chat icon right there because any picture you generate in this app can become a conversation. Let's say we specifically want Nano Banana Pro and Sea Dream to do this. Replace the sea lion with a corgi. And now we have a chat interface for follow ups.
08:17Of course, image generation models still aren't the fastest thing, but they're more than fast enough. And here we are. We made a new generation with a corgi. Still one click to download if you wanna do that. We can zoom in and see him if we wanna get better look. Not bad. How's this one? I like this one more. So we'll pick the Nano Banana one, and we will do one more follow-up. In this follow-up, I will tell it to, I don't know, change the text to say
08:42Corgeco. I'll be more specific. On the sign. This has been a super fun side project both for playing around with Cloud Code and trying to push its limits, but also for playing around with new UX flows. I wanted to see how it would feel to have a two style view, where we hop between a gallery view and a chat view. I was just really curious how it would feel to use a UI like this. And I'm finding that for the most part, it's really good. There are catches. There are always catches. But I am actually overall quite enjoying this experience.
09:11Especially because, thanks to Convex, not only does it sync fully, it syncs fully with the mobile app. I wanna show you guys some of the most absurd prompts I've run-in my life, because I did not think this would work. Here is a real prompt that I sent to Claude Code.
09:28I wanna turn this project into a mono repo with the current web app and a React Native mobile application. For the mobile app, use the following stack. Expo, React Native, Expo Router, TypeScript, Bun, Uniwind for tailwind bindings, and Convex using existing Convex bindings. We only care about iOS for now, so don't put meaningful effort into multi platform support. Ideally,
09:50the convex bindings and definitions will be shared between projects. Use TurboRepo for managing the mono repo and sub packages. Write a thorough plan. This was meant to be a task it could not complete. I got here because I am now convinced there are very few tasks
10:08Opus in the right harness cannot complete. I mostly did this as a joke. I was planning on taking a screenshot of this prompt and posting it on Twitter saying, l o l, nothing can possibly do this. And then it did. Yeah.
10:23This run took over an hour, if I recall correctly. The plan took, like, twenty minutes by itself. Can I go to this plan? Look at that. The plan is still saved. This is just the
10:37mobile app feature implementation. Is there a or maybe it, like, overrode this at some point? Yeah. I think it overrode this at some point. So the history here is the right one. Convert the current web into a Turbo repo monorepo with a new Expo React Native mobile app sharing the Convex back end. You have apps, web mobile, packages, back end shared, TypeScript config for sharing TypeScript config stuff between things, turbo JSON, package JSON, and pmrc.
11:02It had names for all the new packages, designed the foundation with a root workspace config, and pmrc in order to handle the weirdness that is how packages are linked between things in node and mono repos, created the different packages that needed to, moved things over to app slash web,
11:20even calls out critical files it's gonna have to modify and change. Phase two, it extracted the shared packages. You get the idea. I wrote a very thorough plan. I gave it a really quick once over. I don't even think I told it to change anything else, if I recall. I think I just said do it. Looks from here like that's exactly what happened. And then it just ran.
11:42And it ran. And it ran. It even remembered to use its front end design skill, which, by the way, is just a markdown file saying don't make it look like AI Slop. It tried running bun install, it got errors,
11:55it went and fixed a bunch of the versions, It updated the package JSON accordingly. It still got errors. It went and fixed more things. Then it stopped getting errors when it tried to build. It moved from Uniwind to Nativewind at some point because it was having problems with Uniwind.
12:11But it just kept going. Noticed more mistakes with how it was calling the APIs. I need to fix the mobile app to use the correct API functions. Let me update the home screen. And then suddenly, it was done. The mono repo conversion is complete. Here's a summary of what was done.
12:26Yeah. Yeah. To be fair, the mobile repo had a good number of problems. Specifically,
12:34the convex URL was not loading into the mobile app properly. It also did a bad job of copying over my environment variables, both because I had changed to a mono repo and to a workspace, which is the JJ equivalent of a work tree, at the same time as I did this. So I just kinda lost track of environment variables, did a tiny bit of moving files around in Finder to fix that.
12:55Then I kept hitting weird server side errors that apparently were from native wind. Apparently, messages are missing from the history. They're not really great about history with this stuff. So I'm just gonna show you guys the PR. Here is the mobile mono repo PR. The thing that was craziest to me was this. GREPTILE did a review of this PR and gave it a five out of five confidence score the first try,
13:19which I just didn't think made any sense. Called out some places where there were duplicate imports but didn't really matter. And this is a huge PR. That's 2,300 lines added and 400 removed.
13:33It was able to break up one project into five packages with a somewhat complex turbo repo setup with no issues and add a mobile app. There was a slight problem with it adding the mobile app though, which is that it ran the expo init command, which by default includes git, which caused a sub git tree to exist and JJ just ignored it. So I actually forgot to include the mobile app in this when I merged it,
13:57Which shows you how thoroughly I was reviewing the code. I didn't even notice that the app was missing. This ended up bothering me much later as I was trying to do other things with WorkTrees and the mobile app was missing when I did it. Again, turns out I had forgotten that. Ended up asking Claude for help with JJ, fixed my history because again, JJ just doesn't give a fuck. You can go edit your history. It's so nice.
14:19And then re added the mobile app on a separate branch, asked Scriptile to review again, and again, got a perfect fucking score. My hand rolled changes that are like 200 lines of code usually get like a three out of five. And I was hitting fives out of fives constantly with the changes that Opus was making. Now, is GREPTILE going to be the objective perfect
14:39way to review code? It's saying five here and three there means that this is better than that? No. But it is fucking fascinating to see. And if you look through the code,
14:50that all makes sense. Maybe a thousand line of code TSX file isn't the greatest thing in the world, but welcome to React dev. We have all the mutations for all the things that we are doing in a thread.
15:04A timeout that will help with the scroll to end, that all makes sense. My god, the GitHub UI is obnoxious.
15:12It didn't quite get the keyboard avoiding view properly, but it wasn't too far off either. But honestly, I can just show you guys. I should use another mobile phone for this, but I'm dumb. So let's just do it.
15:27And here we are, the mobile version. MacOS's mobile share thing is really annoying with resolutions, but we have the mobile app and it works. And since, again, we were using Convex, everything stays in sync. So if I go and create a new gen here, two corgis
15:46sharing an ice cream cone, and I hit gen here, it syncs with the web UI as well, with no special anything. I didn't tell the model to do this. This is just Convex, which is another one of those things that has fundamentally changed how I build. But one of the things that makes all of this so powerful is that Convex is just a folder in the code base that you configure,
16:08which means the models that need some crazy MCP bullshit to go change settings in a dashboard, it just edits the folder. You get the idea, though. This is a full functioning mobile app that's a lot less laggy on my phone. It's actually been pleasant. Like, I've been using this to do random image gen for some work stuff. The hardest parts of this project by far
16:28were dealing with GCP and the Google Cloud dashboard to get the tokens I needed to do off, and then configuring Clerk, Convex, and all of the Vercel stuff to handle build and deployment properly for a productionized version. The thing that was hard for this was getting it into prod because I had to fight all of these legacy dashboards and tools and things. Everything else
16:52was super smooth. Even other annoying bullshit like implementing auth, which even with the best tools, even with things like Clarke and WorkOS, is not the easiest thing, especially when you have multiple platforms and three different packages in your code base that need to have auth added. I need it in the mobile app, I need it in the web app, and I need it in Convex in a way that convex functions know the user is off and they know who the user is. All of these layers are annoying to get right, even on a good day, and it just did it. It's another 1,800 lines of code added,
17:25but it did it and got another five out of five confidence score. So I merged it. Would I put this code in front of all of our users for t three chat without a more thorough audit? Absolutely not. But is it worth talking about the fact that I was able to, from scratch,
17:41build all of this without opening an IDE a single fucking time? I think that's worth talking about. That's 11,900 lines of code. That's a real code base. That's not a massive code base, but that's a that's a real code base. And this entire thing was generated on the $200 tier of Claude code.
18:01That's just absurd if you think about it. This isn't just tab complete in our editors. This isn't just like, go change these three files for me or deprecate this thing. This is a .
18:17And also a bit of a change on how I think about using my computer. We'll get to that in a sec. But I wanna show one other project I built using all of this. One of the problems I have with Cloud Code and tools like it is now that I'm running these jobs that can take forty five minutes to multiple hours, I find myself getting distracted. And then I don't go back when the work is done, and I don't wanna set up some audio queue or some shit. Even if I did, I don't have the discipline. I'm just gonna browse Twitter, which is why I made an extension.
18:44That locks me out of Twitter unless I have work going in one of my Cloud Code terminals. I'll ask it to make the UI look better, just some bullshit busy work. And now my Twitter's unlocked, and I can once again see my tweet where I complain to the Cloud Code team that I can't hide my email address. But as soon as the work is done, I will be locked out of Twitter again.
19:04This has been a weirdly powerful boost to my productivity. And I have also not opened up an IDE to look at the code for this project at any point. Do I look at it on GitHub sometimes?
19:16A little. Not much. Would I trust this code with my life? No. Not at all. But is it a fun side project that I had an idea for that I was able to build in half an hour instead of a day? Yeah. That's what's cool here. It's changing whether or not I'm willing to make a project, not whether or not I'm capable. I was always capable of building things like this. I'm not doing things I couldn't do before. I'm doing things that I didn't bother doing before. Because suddenly,
19:43they are so much easier to do. Who wish UI service do I wanna improve? Never mind. I like it as is. Nah. And now it's done, and I'm back to locked out of Twitter. That's cool. The fact that I hacked this together in, thirty minutes of work where I was doing other things and working on the mobile app at the same time, that's really cool. I had two of these Claude code instances running for the mobile app, and then one running for this project all at the same time, And would constantly get kicked out of Twitter when I went to talk about it and ask people questions, because when all of the jobs ended, I would get booted. This was such a fun little thing to build. This one's open source, by the way.
20:22Claude Blocker is live on GitHub if you want to play with it. I've already submitted it to the Chrome web Store so we can get the extension out ASAP. Those take a while to get reviewed, especially when you have more grandiose permissions because you need access to tabs to do things to pages. But I could do it. I could do it without much effort. I put this whole thing together without much time spent.
20:41People are already making changes to it, which is really cool. But man, like, this type of thing was stuff that I would maybe do if I had a lot of free time sometimes. And I really wanna emphasize this point, like, these tools aren't just good for making your job simpler or trying to like do work that is tickets from Jira or whatever. The thing that's really fun for me with AI code is that there are all of these ideas and vague things that I like could have done. Like I could have built an app that was just to parse my Notion history, which I did today. I have a different file in Notion for every month where I log every day what I did that day. And I wanted it all to be in one file so I could hand it to LMs to make them judge me and roast me. Just getting all of the content out of those in order was annoying. So I put them in a directory, I opened Cloud Code, I said make these one file, and it wrote scripts to run and make the changes to make it one file. These are all things I could do before. I knew how to write all of this code. But if I was going to spend certain amount of time in a day writing code, it was more important that that time went to projects that mattered, and I didn't have as much left over to do fun side projects. And if I did, I would pick the single biggest most fun side project,
21:50not dozens of them. Now I find myself spinning up random projects all of the time. And sometimes I find myself using Cloud Code for things that aren't even projects. Remember that thing I said earlier about the way I use my computer change? Let's just look at some examples. I need you to update my JJ config to sign commits the same way that my git config does. This one, I think I had multiple messages for.
22:14I guess it's just not restoring properly. That is really annoying. It seems like resume is just kinda broken. Whatever. The thing I wanted to show is that I was able to update my JJ config to fix my commit signing by asking it to, and then it did. I was able to add a new script to my z shell that was a script that will automatically break something out into a work tree, change to that directory,
22:39and copy over all of my ENV files from that project by just asking it to do that. I'm not just asking Claude Code to edit files in a code base and occasionally search the web for more context. I'm asking it to use my computer and make changes to my setup in my environment the way I normally would between like five different tabs, a bunch of searching, a bunch of trial and error, or I can just tell Claude Code to do it and go grab a tea. It's unbelievable.
23:05There are so many things that just didn't make sense for me before, be it a new side project, be it a weird extension I wanna build, be it configuration changes I wanna make to my system, be it the move to JJ instead of Git, where I'm starting to move off of Git as my default, like, CLI and way of managing repositories, these types of things weren't as viable before,
23:27and now with things like ClaudeCode, they are suddenly much more viable. If you're still new to tools like this, I won't necessarily recommend that you do what I'm about to show you immediately. Definitely keep it in the standard prompt you for edits mode, then when you have the confidence, switch over to auto accept edits. And then, after you've done this for a while and you've let it run commands that edit your system, change things, and take risks maybe a little more than you expected, and were surprised with the results, like I and many others have been, you can switch over to allow dangerously.
23:57I understand why they hide it. I understand why they are so strict about this. But allow dangerously skip permissions, it's so fun. It's genuinely so fun. I admittedly have a bias here, which is that if I end up being one of those people where Claude randomly nukes my home folder,
24:14I get really good content out of that. If that happens to you, you probably don't have benefits. It probably just sucks. If that happens to me, it'd be one of the greatest days of my life. I could make so much good content about that. But it probably won't. But it could. But I'm also doing a few things to prevent it, because I wanna do my best. Due to a series of unfortunate events on Twitter I do not feel like going into, I was made aware of this new project, the Claude Code Safety Net.
24:40This is a plugin that you can add to ClaudeCode. Shout out Daisy for the plugin system. It's awesome, one of the coolest things about it. It's a ClaudeCode plugin that acts as a safety net, catching destructive git and file system commands before they execute. So if the model decides to do some crazy shit with git or some crazy shit with rm, it will prevent it from doing that and give you one last
24:58check to the user even when it's in the dangerously modes. Will this catch everything? Absolutely not. You could still write a bash file and then put it in your system and then execute it that does those things, and it can and it will, the models love to try and work around the restrictions they have. If they can't edit a file by using the edit tool, they will run a bash script to edit it instead. If they can't do that, they'll run a Perl script to do it instead. The models are very, very
25:25weirdly willing to work around the restrictions you give them. So know that this will never be perfect. But again, for my risk profile and the things that I'm concerned about, the combination of the Cloud Code Safety Net plugin and my own traditional usage of ClaudeCode in AllowDangerously has not come close to causing me any problems.
25:44I've been impressed. This has been a wild deep dive over the last two weeks, and I don't know how it's going to affect the way I code long term. Obviously, I still love Cursor. I still have my investments there, so account for the biases accordingly. When I actually want to see the code and do my job as an engineer, not just dick around with a lot of different things and ideas, I still very much like using a tool such as Cursor. And when the code base already exists and has things going on in it, and I'm working in a code base as my job, Cursor's whole agentic environment is still something I really enjoy. But when I'm experimenting, when I'm screwing with my system, when I am in some crazy greenfield thing,
26:25HOOKI just wanna let something run-in the background for an hour on some crazy task, I've been impressed with Cloud Code. Now, there's two questions I know are already swarming in the comments section because the people who are asking these probably didn't watch till the end. Sorry to them. Question one. What about OpenCode? OpenCode's really good. Not gonna pretend otherwise. I wanted to try out the new Cloud Code updates in the new Cloud Code subscription tier in the way that a lot of people have been using it, because I haven't really had the Cloud Code clicked for me moment yet. And this is what pushed me over the edge. It's a combination of Opus 4.5 being an unbelievably smart model,
26:57HOOKand the harness getting mature enough that it's in a really good spot. OpenCode is still the best option for almost every model. And it probably is almost as good when using Opus. And they also have a way that you could use your Cloud Code subscription, and Open Code is a bit hacky, but it seems to work fine. I just didn't feel like dealing with any of that. I'll evaluate it more in the near future. And now the next question, which is the Ralph Loop. This website is questionable. I'm assuming it was made by a model that was not given the
27:25front end design skill. That said, if you're not familiar with Ralph Wiggum, he's a character from Simpsons. Also, hi, Gen Z. Good to see you guys. I know everybody else is familiar with The Simpsons. The point of the Ralph Wiggum loop is to let Claude code continuously keep working even when it thinks it's done until some higher order, this is probably done, gets hit. Also, Miles decided to come say hi. Now that I am doing more vibe coding, I spend a lot more time petting my cat when I'm coding. So he now assumes when I'm working on code
27:57that it's cat petting time. You like Claude code a lot, don't you, buddy? He prefers Codex because it's so much slower, so I get more time to pet him. Aw. He's being so cute.
28:10Anyways, as I was saying, Ralph is a bash loop. This is a way to run Claude code and keep running Claude code even when it stops and asks, are you sure or do you want me to reconsider this or I finished phase one, should I start phase two? Just run-in a bash loop and keep telling it continue, continue, continue until eventually it actually finishes. It's a fun strategy.
28:32Personally, I found that Claude code is pretty willing to just go for an hour or two without being told, it's okay, you can continue. I've had one time total thus far, where it would stop after each phase and I'd be like, no, keep working, please. No, keep working, please, over and over. But only once ever did I hit that, and it was actually a pretty short run. Like, that whole run only took five to ten minutes. But I've had other ones that took an hour and didn't stop at any point. So I still personally don't know how necessary this is. But believe me,
29:01the Ralph loop is something I very much have planned. So if you don't see a video about the Ralph loop from me soon, know that I wasn't impressed enough or didn't find it interesting enough to do one. But if you are interested in this, keep an eye out. I will almost certainly be talking about it. I think that's everything I had to say here. I'm impressed.
29:20Claude Code's gotten a lot better. I have my complaints. Like, the hooks still suck and feel half finished. Plug ins are awesome, but don't have all the functionality I need to be able to rely on them directly. Skills are kind of a joke. They're literally just markdown files. There's a lot of little UX things like the way stashing works sucks. If you wanna change things after you wrote your prompt, you're screwed.
29:41Context compaction feels really strange still. The way history management works and the way picture uploads work is still jank as hell. There's lots of little places it can improve. I don't want this to seem like it's just this perfect magical solution. But goddamn, I have been able to do a lot with this. And I've been really impressed.
29:59CTAAnd one actual last thing, my usage. Remember, I was on the $200 a month tier with the two x limits. I did manage to hit the limit on the $20 tier. I considered going to the five x, which is $50.01, and then up to the $200.20
30:15CTAx one. But I decided to just go all the way and see how much I could use. Like, I was going out of my way. I was running two or three of these in parallel during most of my waking hours. I want you all to guest chat mostly between like 1390%, some people with some jokes. You can't really know your usage and they work really hard to hide it. But if we look at the dashboard,
30:38CTAyou'll see that my weekly limit, I'm at about 2% of my usage. So this just reset. So that one being 2% makes sense. But this top level one, my plan usage, which resets
30:50CTAthirty four minutes for this session, 12%. And the highest I managed to get in my weekly limit last week was 7%. 7% going as hard as I could. Take off that two x and I was 14%
31:06for the week. From my napkin math from the tokens I was able to log, I did about $1,500 of inference in that time.
31:15I paid $200 and I got 1,500 out. As a user, I love this. That is a really good deal. As a company also selling inference that pays API pricing for a lot of Anthropix APIs, I am pissed that I'm subsidizing this.
31:31The reason these subscriptions can give you 10 x the inference you pay for is because they are subsidized by people like me paying full price over API. So on one hand, cool. I love using this. On the other, you're welcome, you ungrateful bastards.
31:47God. So much clicked over the last few weeks for me. I I have a whole different world view, a whole different view on how code works and where it's all going, and why people like Claude code so much.
32:02I got nothing else. I hope you enjoyed this, this descent into chaos.
32:09Let me know what you guys think, and if Claude Code is driving you mad as well. Hope this was useful. But until next time, peace nerds.
— full transcript
§ 05 · For Joe
The threshold is viability, not capability.
Builder playbook
Claude Code does not give you new abilities — it makes the things you already could do worth doing.
- Use a front-end design skill markdown file in every project to prevent AI-slop UI without extra prompting.
- Spin up 3 routes for the same feature, compare visually, cherry-pick — faster than iterating on one.
- Convex as just a folder makes models self-sufficient on backend config — no MCP dashboard required.
- Add Claude Code Safety Net before flipping to --allow-dangerously. One plugin, big risk reduction.
- Greptile as a proxy reviewer: Theo's hand-rolled code got 3/5, Claude's got 5/5 consistently.
- The Ralph Wiggum Loop feeds continue in bash — rarely needed with Opus but useful for multi-phase tasks.
- The $200/mo tier gave ~$1,500 in inference. API customers subsidize subscribers — wild pricing math to understand.
§ 05 · For You
What this means if you want to start using AI coding tools.
For the curious builder
The real unlock is not the quality of the code — it's the lowered threshold for starting things you already wanted to build.
- Start in the default prompt-you-for-edits mode before switching to auto-accept.
- These tools work dramatically better if you already know how to code — you need to verify the output.
- Side projects you never started because they were not worth a full day now take 30 minutes.
- You can use Claude Code for system config changes not just code projects.
- The $200/month tier is a wild deal if you are actively building — Theo used ~$1,500 compute for $200.
