CLAUDE.md Is Guidance, Not Verification
Not everything should be automated. Map each pull request to a lane, then see why CLAUDE.md only guides and what actually verifies the work: rules, gates and hooks.
Transcript
Does that mean software factory means everything needs to be automated? Definitely not. I myself have 14 years of software engineering experience, and I have been a software engineer and I lead a software engineering team.So I'm not someone who says like you can automate each and everything without human. It is not the case. Software factory doesn't mean you need to automate each and everything. Rather, you need to set a process that way, whether I produce it or my colleague produce the output for that ticket. The output is consistent enough and you are able to verify that output with 100% confidence. That is what a software factory means. How much automation is purely depends on your company's immunity? So, for example, when I'm building out land with param platform, it's a very simple community platform where I teach, and you know, in this case, even if it breaks somewhere, nobody is going to care because it's not a financial institution or something where they are losing money. Of course, they might be signing in and then they come and say, like param something is failing here. you can fix it, and I will go and fix it. So that's purely immune enough, and I can be able to build it. In a in a in my company website, like cast.team, that is okay. But if you are operating in a financial institute, your immunity is very different. You can't just merge it and after that, you will find out, okay, something I need to fix it. And once again, I will come back and fix that. That you need to create that plan, like what is high risk, what is low risk, what which ones of these PR I can review, which one requires human review, which one requires pair programming, which one requires you know, fully autonomous run and so on. That is something you need to map in the software factory process or in the mindset itself, like as an engineering process, and then you need to execute what are all the pieces we can automate and how the events can be connected to make it semi-autonomous or fully autonomous process from issue to PR. Okay, as I said, whether it is Claude code or any of the coding harness, if you are just saying the guidance, so for example, most of the time CLAUDE.md or agents.md is just a guidance, so you do this, you don't do this, and so on. But does those coding agents always follow that guidance? No, definitely not because maybe in a small repository, yes, but in a larger repository where the context is beyond one million of something you can fit in inside the context memory, then most probably it won't. it it it will start to hallucinate or it will start to find different ways to solve the problem, and while one or two instances, it can definitely fail on such things, you can't just feed hundreds of rules in one single markdown file and fit into the memory. So that is where like you need to have a clear deterministic way to handle how you you are going to get and how you are going to test it, and how you are going to you know deny some of the things that shouldn't be happening inside your repository that you need to set with Claude. I suggest to read one of the articles that is called Claude Directory from Claude Code. again, it it is very specific to Claude Code, but it the same principle can be applied for any of those. So Rules are the cheapest and also the weakest. We can set a rule, but doesn't mean rules will be always executed. You need to go with gates and test check with no model so that it will be more deterministic in nature, and you need to have a hook to refuse a tool call before it runs, so that you can you know, Pre-hook, post hook, you can always classify what it can read, what it can write, and so on. And stop hook, whenever you think something is going against or something is not going in the right direction, the stop hook can stop it as well, and then you can have a dedicated reviewer to review the diff.
This lesson comes with Pro, $149 a year: Get Pro. This lesson is cut from the free Lightning session Build a Claude Code Verification Harness. The Live Labs build the same system with you, free, on Thursdays: see the Live Labs.
