This project supports multiple LLM coding assistants through universal context files.
| File | Purpose | Consumers |
|---|---|---|
AGENTS.md |
Primary agent context | All LLM tools |
CONVENTIONS.md |
Code style guide | All LLM tools |
specs/*.feature |
BDD behavior specs | All LLM tools |
specs/*.schema |
Type definitions | Generators + LLMs |
specs/*.sm |
State machines | Generators + LLMs |
All symlink to AGENTS.md (single source of truth):
| File | Provider |
|---|---|
.claude/CLAUDE.md |
Claude Code (Anthropic) |
.cursorrules |
Cursor |
.github/copilot-instructions.md |
GitHub Copilot |
AI.md |
Generic / Aider |
LLM.md |
Generic |
CONTEXT.md |
Generic |
AGENTS.md ← canonical source
↑
├── .claude/CLAUDE.md
├── .cursorrules
├── .github/copilot-instructions.md
├── AI.md
├── LLM.md
└── CONTEXT.md
LLM tools should look for context in this order:
AGENTS.md(universal, preferred)AI_CONTEXT.md(this file, meta)CONVENTIONS.md(style guide)README.md(project overview)- Provider-specific files (
.claude/,.cursorrules, etc.)
- Portable: Works across IDEs and LLM providers
- Maintainable: Single source of truth
- Discoverable: Standard filename
- Human-readable: Also serves as documentation
Following Cosmopolitan's philosophy: write once, run everywhere. Same principle applied to LLM context: write once, understood everywhere.