GuidesWrite Playbook docs

Guides

Extract durable agent rules

Keep the source as a normal Topic and mark only the durable rule, closeout, or scoped instruction that agents need.

Agent instructions

Mark only the rule agents need.

When someone says “agents should”, “make Codex always”, or “Claude should remember”, keep the guidance in a Topic and wrap the durable instruction in an agents-md marker.

Use lifecycle only when a script must run automatically. Use team-member for a delegated specialist role, command for a user-invoked procedure, and workflow for multi-agent orchestration.

docs/standards/agent-operating-contract/overview.mdmd
---
name: agent-operating-contract
title: Agent Operating Contract
description: "Instruction fidelity, scope control, verification, and proactive boundaries for AI coding agents."
---

# Agent Operating Contract

<!-- agents-md:rule scope=. order=10 id=instruction-fidelity -->
Follow the user's requested scope exactly. Do not broaden the task, redesign adjacent surfaces, or refactor unrelated code unless explicitly asked.
<!-- /agents-md:rule -->

<!-- agents-md:rule scope=. order=20 id=proactivity-boundary -->
Be proactive inside the requested goal and conservative outside it. Mention adjacent improvements as options instead of implementing them silently.
<!-- /agents-md:rule -->

<!-- agents-md:closeout scope=. order=10 id=verified-closeout -->
Before saying done, verify the real target surface, report validation evidence, and call out anything unverified or risky.
<!-- /agents-md:closeout -->
How markers route

rule renders into hard rules, closeout renders into handoff requirements, scope=. targets root project instructions, and order keeps output stable.

Why not edit generated files?

Generated AI Channel files are outputs. Change the Playbook source, validate, dry-run sync, then write the generated files.

Guide map

Choose the decision you are making.