Agent Playbooks

Every tool wants its own file. Claude Code, Cursor, Codex, and Copilot each expect instructions in their own format and location. agent-playbooks is one generic instruction set instead — write the rules once, point every tool at the same source of truth.

An agent left to its own judgment fails in the same few ways, every time. It guesses at an ambiguous spec instead of asking. It grades its own work instead of checking it independently. It forgets a project's own conventions the moment a new session starts. It occasionally runs something destructive because nothing stopped it.

This is that process, written down once, usable everywhere. Every playbook has been run against a real, throwaway test case at least once — not just written and published.

Every release is signed. An Ed25519 key that never leaves the maintainer's machine signs a manifest of every file's sha256 hash. install.sh verifies the signature and re-hashes every fetched file before writing anything — a spoofed backend can no longer silently swap in different content.

A real, enforced block on destructive commands. Not a markdown instruction hoping the agent listens — a deterministic script wired into the tool's own pre-execution hook where supported, with its exact coverage and gaps documented, not assumed.

Nothing hidden about what gets sent. The installer sends one random local ID, nothing else. No account, no telemetry beyond that, no email. Read supabase/functions/check-in/index.ts yourself — it's the whole mechanism.