Most people write AI instructions as a long system prompt: a wall of text at the start of every conversation. This works, but it has limits. The instructions are not versioned, not reusable across contexts, and tend to grow bloated over time.
The Skill File Approach
We store AI instructions as individual markdown files, one per skill or task type. A file called commit.md contains the steps for creating a git commit. A file called blog-post.md contains the brief for writing a post in our style. Each file has a clear purpose, a version in the frontmatter, and gets loaded into context only when needed.
Why This Works
You can update a skill file without touching every prompt. You can test different versions. You can reuse the same skill across different projects. And because the instructions are in plain markdown, a human can read and edit them directly, no special tooling required.
The CLAUDE.md Pattern
Anthropic's Claude Code tool uses a CLAUDE.md file for project-level instructions. We have extended this pattern, a global CLAUDE.md for general rules, plus per-project files for context specific to that codebase or client. When Claude starts a session, it reads the relevant files and carries those instructions throughout.