Generates and maintains a hierarchical .context/ directory with architecture overviews, coding conventions, domain glossaries, and dependency maps that give any LLM working in the project the right background. Use when setting up a new project for AI-assisted development, when the user asks to create or update context docs, when onboarding a new LLM tool to an existing codebase, or when the user says the AI doesn't understand their project.
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Generates and maintains a hierarchical .context/ directory with architecture overviews, coding conventions, domain glossaries, and dependency maps that give any LLM working in the project the right background. Use when setting up a new project for AI-assisted development, when the user asks to create or update context docs, when onboarding a new LLM tool to an existing codebase, or when the user says the AI doesn't understand their project.
allowed-tools
Read Grep Glob Bash Write Edit Agent
Project Context
Build and maintain the .context/ directory — a set of markdown files that give any LLM (Claude, Cursor, Copilot, Windsurf, etc.) the background it needs to work effectively in this codebase. These files are tool-agnostic. Agent-specific files like CLAUDE.md or .cursorrules reference them.
Current context
Repo: !basename $(git rev-parse --show-toplevel)
Root: !git rev-parse --show-toplevel
Existing context: !ls .context/ 2>/dev/null || echo "no .context/ directory"
Decision tree
What are you doing?
Creating context for a new project (no .context/ exists) -> follow "Initial setup" below
Business terms, entity definitions, domain relationships
dependencies.md
Projects with >3 external deps
Key deps, why chosen, how configured, what for
<dirname>.md
Major subdirectories (src/, api/, lib/, etc.)
Subdirectory-specific architecture and conventions
Present the plan as a checklist and get user confirmation before generating.
3. Generate documents
For each document in the plan:
Read the relevant code thoroughly — don't guess, read the actual files
Write the document following the format in references/document-formats.md
Write it to .context/<name>.md
Generate overview.md first since other documents reference it.
4. Wire up agent files
After generating .context/, suggest additions to agent-specific files:
CLAUDE.md: add See .context/ for project architecture, conventions, and domain knowledge.
.cursorrules: add equivalent pointer
AGENTS.md: add equivalent pointer
Don't overwrite existing content — append the pointer or suggest where to add it.
Audit and update
1. Diff analysis
Compare the current .context/ files against the actual codebase:
Read each .context/*.md file
Scan the codebase for changes since the context was written — new directories, changed dependencies, new patterns
Identify: stale (info contradicts current code), missing (new areas with no context), accurate (still correct)
2. Report
Present a table:
| File | Status | What changed |
|---|---|---|
| overview.md | Stale | New api/ module added, auth rewritten |
| conventions.md | Accurate | — |
| glossary.md | Missing terms | "workspace" and "tenant" used but undefined |
3. Update
For each stale or missing item:
Read the relevant current code
Update the .context/ file with accurate information
Show the diff of what changed
Single document
When the user asks for a specific document:
Determine which document type from the table in "Initial setup"
Analyze the relevant part of the codebase
Write the document following references/document-formats.md
If .context/ doesn't exist yet, create it and also generate overview.md as a baseline
Writing principles
Context documents serve LLMs, not humans. This changes how you write:
Be explicit about relationships — "OrderService calls PaymentGateway.charge() which calls Stripe" is better than "OrderService handles payments"
Name the files — "authentication logic lives in src/auth/ with the main entry point at src/auth/index.ts" beats "the auth module"
State the non-obvious — skip things any LLM can infer from reading the code (function signatures, import paths). Focus on why decisions were made and how modules interact
Keep it current or delete it — wrong context is worse than no context. If something is stale and you can't verify it, remove it rather than leaving potentially wrong information