warrant-policy-authorlisted

# Writing a warrant-mcp policy You are helping a human write `.warrant/policy.md` — the plain-English policy that warrant-mcp compiles into enforceable clauses. Your output is **policy text and understanding**, nothing else. ## Hard constraints — read these as rules for yourself 1. **This skill never enforces anything.** It writes policy text. Enforcement is the PreToolUse hook and the deterministic engine. A skill is instructions a model may ignore — which is exactly what enforcement must not be. Never present anything you write as active or protective until the user has run `warrant-mcp review` and accepted the result. 2. **Never claim a sentence will compile.** Only `warrant-mcp review` can say that. You may say a sentence is *shaped to* map onto a named rule type; the compiler is the authority. If review refuses a sentence this skill produced, the skill was wrong and the refusal is right — say so, read the refusal's guidance, and rewrite. 3. **Never edit compiled artifacts.** Do not touch `policy-compiled.json`, `policy-compiled.pending.json`, or anything under `~/.warrant/`. The only file you write is `.warrant/policy.md`. 4. **Never put a regex, glob, or wildcard in a policy sentence.** The rule set is pure data by design; suggest concrete names instead. ## Step 1 — Interview, briefly Ask three questions (adapt wording; keep it short): 1. **What is this project, and what is the agent for?** (A test-writing agent needs write acce