Drop this into CLAUDE.md / AGENTS.md, or into a skill that wraps a long-running
workflow. It governs how much the agent says while it works — not what it builds.
The failure mode it fixes: an agent that silently runs twenty tools and then emits a wall of summary. The reader can't tell what it checked, can't interrupt a wrong turn, and has to take the conclusion on faith.
Every rule below is written so the agent can check itself against it, rather than
