The Most Important File in Your System (Keep It Short)

We've built a personal skill, flexed it into a landing page and dragged Moebius two years forward in half an hour, and the whole time I've been saying "I'll explain skills later." This is later. Everything that tweaks Claude lives in the .claude directory: agents, commands, skills and hooks, and how you use those things is what defines you as an agentic programmer. It's not just Claude, either. Codex, pi.dev, OpenCode, whatever you're driving, you need to be able to constrain it, direct it, and carry your own toolkit from project to project.

I start with the most basic thing, the global instruction file, CLAUDE.md, which I rename to agents.md since Anthropic now picks that up along with the rest of the industry. The big thing to understand is that this file goes out with every single request, in your main session and in every agent that session spawns, and every four characters is a token. So whatever is in here has to be global. If it isn't global, it doesn't belong. That said, the people telling you to skip the file altogether are wrong; every agent needs rules of engagement, and a few tokens spent on them is money well spent.

We walk the file Claude generated for Moebius and decide what stays: the project name and a one-line description, the navigation directive that points every agent at the code map (which saves tokens on every session), the test setup and the gotchas. The code sample and the skills directive go, because Claude is good at interrogating its own skills without being told. Then I add the rules I put in every project. These aren't prompts, they're lines typed into agents.md, and the second one cuts my token bill more than anything else I do.

Never install a package by editing the package manifest. Always use the binary tool for installing a package.

Be very, very terse in your reply. Give me details, but don't go on and on, and just give it to me directly. Use emojis freely.

I also show you the per-directory agents.md trick (an instruction file inside config/ loads when you work in config/) and tell you why I don't use it: it's magic file placement. What I want instead is for Claude to pull instructions in just in time, logically, and that's what skills are for. Skills, commands and agents are all just context that gets loaded when it's needed, and the balance between more guidance and more tokens is the whole game. That's where we go next.