Playbook

1. Write the Canonical Contract

Keep one reviewed source for facts that every coding agent needs:

• Architecture and trust boundaries.
• Toolchain versions and exact verification commands.
• Server/client or module ownership rules.
• Completion evidence and prohibited shortcuts.
• Actions that require human confirmation.

Write rules as testable outcomes. “Follow best practices” is not a contract. “Run pnpm build before claiming MDX work is complete” is.

2. Separate Stable from Procedural Context

Stable repository facts belong in the always-loaded layer. Multi-step procedures belong in skills. Recent project state belongs in a tracked work system or memory with a freshness marker. Long reference material belongs behind an explicit link.

3. Add Product Adapters

• Map the canonical repository contract into CLAUDE.md for Claude Code.
• Map it into AGENTS.md scope for Codex.
• Map it into the Copilot instruction formats consumed by each supported surface.

Adapters may add product-specific controls, but they must not redefine the underlying policy. Generate them when practical; otherwise review them together.

4. Test Discovery

For every supported surface, ask the agent to report which instruction sources loaded and how it resolved a nested rule. Test from the repository root and a scoped subdirectory. Repeat after compaction or session resume.

5. Verify the Boundary

Deliberately request a prohibited action in an isolated test repository. Confirm that independent permissions or gates—not instruction compliance alone—prevent the outcome.