Agent Md — AI skill for Claude Code

Agents.md

**The field guide to AI context files** — the small set of Markdown documents that quietly decide who your agent is, what it knows, what it can do, and how it behaves.

If you've ever dropped a `CLAUDE.md` into a repo, scribbled notes into a `MEMORY.md`, or wondered why some agents feel coherent across sessions while others forget everything the moment you close the terminal — this guide is for you.

These files aren't a framework you install or a spec you have to obey. They're **conventions** — the same way `README.md`, `LICENSE`, and `.gitignore` became conventions long before anyone standardized them. A handful of teams started naming their agent-configuration files the same way, the names stuck, and now they form a loose but powerful vocabulary for *programming agents in plain English*.

This document is a deep, practical tour of that vocabulary.

---

Table of Contents

1. Why context files exist
2. The mental model: a layered identity
3. `CLAUDE.md` — the operating manual
4. `MEMORY.md` — persistent knowledge
5. `SKILLS.md` — the capability registry
6. `SOUL.md` — identity and values
7. How the files compose
8. The wider machinery
9. Anti-patterns and best practices
10. A complete worked example
11. References

---

1. Why context files exist

A large language model is, by default, **stateless and contextless**. Every conversation starts from zero. It doesn't know your coding standards, your name, which tools it's allowed to call, or what it learned yesterday. You *can* re-explain all of that in every prompt — but that's slow, error-prone, and impossib