How the Creator of Claude Code Runs Agents

I use Opus plan. This is literally like secret stuff. It Boris Churnney, he invented core code and I I just want to say that although he invented Clawude code, claw code would not be what it is if it wasn't for the team around him. So this is not to necessarily say that they are the you know what they say goes, but there's absolutely something you will probably learn from this and you might learn a thing or two as will I just kind of going through this. So he runs five claudes in parallel in his terminal. I number my tabs one to five and use system notifications to know when cla code is input. So I'm guessing work trees play a big part to be able to do this unless he's using different git repos for you know they're but I guess this is building clawed code. So I'm get I reckon this is work trees. So I have I have a video on work trees on my own channel if you want to go check that out. But it allows you to run different clawed code um instances all working on different features. So yeah, my experience in this it can get very confusing because you think [ __ ] where was I? What was I doing in that one tab? What you know where was five is quite a lot. I struggled with three put it that way. I also run five to 10 clauds uh on claude.co. So that is claude for web. You can also achieve this with the desktop app. It looks and behaves exactly the same. This is a really nice way to like just send off something and it it creates a pull request on on your GitHub repo. You don't need to run it locally. Um, as I coding my tone, I'll often hand off local sessions to web using at which is something you literally type at and then it will send off to web which is quite cool or manual kickoff sessions in Chrome. Sometimes it will teleport back and forth. So basically he's got up to 15 tasks running in parallel which is quite mental. Mhm. I use Claude uh I use Opus 4.5 with thinking for everything. It's the best coding model I've ever used and even though it's bigger and slower than Sonet since you have the steer less uh it less and it's better tool it is almost always faster than using a smaller model in the end. Now he obviously works for Anthropic. He probably has free use, infinite use of Opus 4.5. This is not practical, but what I do find quite interesting, he's got thinking on for everything. Now, you can enable thinking with um uh option or alt key t and you can turn enable thinking on. I'm in an open question at the moment. What is the difference between thinking and planning? Because you're not doing any both are not doing any code. thinking tends to think before it implements all in the one like uh shot whereas planning gives you the opportunity to you know thumb up or thumb down something I'm discovering but it is just interesting that he's just got thinking on for everything again he works for anthropic this is probably very uh unachievable by normal people our team share a single claude MD file for the claude code repo we check into git and the whole team contributes multiple times a week anytime we see claude do something incorrectly we add So they're actively contributing to this cruel MD file. Um, which is, you know, I think we all should all be doing this. Like you're it's not going to work because it will be unique to your way of working. It will be unique to your team's way of working, but it could also be unique to your um the way that you use cla code. You know, you might you might prompt it in such a way where it always does uh what has he got here? um you know it always uses npm because of the way you write code. It's a stupid example but you know what I mean like your experience with claw code is going to be very different to another person. So your rules are interesting. Yeah. During a code review I'll often tag at claude on my co-workers PRs to add something to the claude MD as part of the PR. We use the claude code GitHub action which you can install using slash whatever um for this. It's our version of Dan Shipper's compounding engineering. I think he's basically just encouraging number four here. Most sessions start in plan mode. There you go. If my goal is to write a pull request, I will use plan mode and go back and forth record until it's I like its plan. From there, I switch into auto accept mode and cause can usually oneshot it. Again, he's probably using Opus, but I use Opus plan, which is a hidden model. You type /mod and then type opus plan and it will unlock. This is literally like secret stuff here. It will unlock whenever you're in plan mode. It will go to opus and then it will switch to implementation mode and automatically switch to sonet to actually implement that stuff. He's rich. He can do opus all the way, but that might be a more cost-effective way to to do this. I use slash commands for every inner loop workflow that I end up doing many times a day. This saves me from repeating prompting and makes it that Claude can use these workflows too. commands are checked into git and live in claw command. I use commit pushpr/comand dozens of times a day. The command uses inline bash to premputee and a few other pieces of info to make the command run quickly and avoid back and forth with the model. Not quite sure on that one. Let me know if you if you understand that. I use a few sub aents regularly. Code simplifier simplifies the code after claud is done working. Verify app has detailed instructions for testing claude code end to end and so on. Similar to slash commands, I think of sub aents as automating the most common workflows that I do. So he's a big sub aent user on call guide. We use post tool hook to I really want to get into hooks. Post tool hook to format Claude codes tool. Claude usually generates well formatted code out the box and the hook handles the last 10% to avoid formatting errors. Okay, so he just runs format basically. Um, I don't use dangerously skip permissions. Instead, I use permissions to pre-allow common bash commands that I know are safe in my environment to avoid unnecessary permission prompts. Most of these are checked into Google. So, okay. So, yeah. Um, you can basically set a bunch of uh permissions that it's allowed to run instead of giving it basically you can do anything. He's being a bit more controlled about what it can run while it's doing its thing. Um, claw code uses all my tools for me. It often searches and posts to Slack via the MCP server. Runs BigQuery commands to answer analytic questions using B big BQ CLI. Don't know what that is. Grabs our logs from Sentry. Uh, the Slack MCP configuration is checked into our MCP JSON and shared with the team. I think he's basically saying that he just does everything from SL uh claw code. I think via MCP for very long running tasks. I will either a prompt claw to verify its work with a background task when it's done. This is something I want to talk about. Use an agent stop hook to do that more deterministically or use the Ralph Wigum plugin originally dreamt up by Jeff Jeffrey Huntley. I will use either permission mode don't ask or well he just said he doesn't use that in a sandbox to avoid permission prompts in a session so Claude can cook without being blocked by me. you kind of mentioned it a little bit earlier, this idea of I think what would be more common is kind of what uh PewDiePie built where you've got a council of agents all verifying each other's work. I think this might be what he's kind of doing here like creating a sub agent to verify others work but not to 100% sure. I ask my AI like I always um give it like I ask it to rate the code and sometimes I do it like manually with other AI and what I've recently started doing is I find like either open source or like yeah I find code that I think uh would make sense to kind of like um compare it to. So I say like compare this what you created with this and then write it. And then uh I also started doing like doing that and putting the end results in an MD file for either for AI to learn or for myself maybe both of us like so they learn from the the code that they did or sometimes you know when you get an error or like you get a bug or something or something doesn't work out and then you continue prompting until you get it fixed. Uh I say everything that you learned through like this back and forth documented in an MD file. Why was it so and how is that that it's it worked. Uh so I do try to gather these either for myself to learn from or you know I could feed it back to uh AI in another session if I have something similar. I don't know if you've heard of Ralph Wickham. It's doing a lot of rounds at the moment, but it was um it's basically a task that runs infinitely until the job is done. So, and it feeds the result back in. It's just an infinite loop basically. You can set up max iterations, but iterations. But if you look at cold code, look at plugins, the plug-in directory, you'll see Ralph Wiggum. Um I made a short about it actually, but um yeah, it's becoming more popular like in the last week. So final tip, probably the most important thing to get great results out of Claude code, give Claude a way to verify its work. As you were just saying, if Claude has that feedback loop, it will be two to three uh times the quality of the final result. So there we go. Again, everyone's going to have their own ways of doing things. Everyone's going to enjoy different ways of working. That might not work for you, but certainly works for them. Can't argue with it. I wish I could, you know, obviously literally just show you. I mean, I probably could, but it would just make this section like 10 times more uh longer if I'm there showing you exactly how to do it. But I hope that's interesting. That that's the idea for the comment section. Like just let us know if you would be interested in that. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.