The Repo Carries the Harness

let me show a bit of demo today. this is a repo, it's a small repo. I have created mainly for today's session and software factory. This repo have you know it's a small bill splitting app, nothing on my you know nothing fancier in that sense, and it has a Claude repository. And it has Claude directory with certain settings, rules, and skills along with the agents and hooks, just to make sure this demo is valid enough. And if you see the rule part, it loads only this rule is more related to the money code, and it gives context about how you can you know solve any money-related details. And in agents.md, this rules will be called. So automatically, this rule will be applied for whenever there is some code changes which happens, and then a skill. this skill is very exclusive for this repository. A skill to fix an issue. What you need to do to fix an issue, and it will basically call these are the steps you need to follow in order to solve an issue. does that mean always you need to create such such skill? No, it is not again. Processes you don't need to create inside Claude Code in most cases. but just for the demonstration purpose, I created this skill and it will be following based on this skill. And a sub-agent, that sub-agent is basically a second reviewer with its own context, so it just judges the differences. like as I said, you don't need to use your agent alone to test what it codes because they have their own context and they build based on that context and they will verify only the things they changed. So oftentimes every agent will say, I am done. So you need to have a sub-agent for your coding reviewer. There are companies which even does you know, not sub-agents, but but rather they do like okay, I will do Opus for planning like sonnet for actual execution, and then I will call Codex or Code Rabbit in CI or in headless environment in serverless environment for code review for the different diff in place, and that will do a review and then post okay, these are the changes which is not right, and these are the changes I'm having problem, and it will comment on your GitHub PR and you need to resolve the PR, and then once again you need to feed in. There are different ways companies are automating, depends on what are all the what are all the tools they like and what are all the ways they wanted to do it in their team. in this demo, I created a simple claude agent itself and all these things as I said you can verify these inside over here. So agent hook rules skills and headless JSON settings.json, everything is already there. So you you you you can go and check it out on the repository actually. Okay, and as you see, most of the time you you you say what are all the things you wanted to deny. This is default Claude rules. you can create settings.json or you can create settings.local.json. In some cases, you will create your own rules, like okay, deny all these, only allow these rules or these calling and tools.