review-plan-risklisted

# Review Plan Risk Find and fix **design risks in a plan before it is built**. Humans and LLMs share a positive test strategy: reading a coherent plan pulls toward confirming it. This skill replaces "is this good?" with deliberate disconfirmation — search for the ways the design fails, then close them in the plan. **Scope: the design, not its execution.** If the target is code, a diff, or an existing implementation, say so and stop. Runtime concerns (crashes, races, malformed inputs) belong to implementation review — unless the design fails to account for them, in which case the gap is the risk. ## Step 1 — Scan the plan Read the artifact in full — a plan, spec, RFC, or a workflow definition (SKILL.md, command, agent prompt; a first-class case). If no concrete artifact is given, ask for one — do not pre-mortem a vague idea. A caller may additionally name a **baseline copy** of the artifact in its invocation — a pre-fix snapshot of a file it deliberately keeps untracked and owns; note it if present, as Step 3 uses it to decide whether an otherwise-unrecoverable file can still be auto-fixed. The artifact may span more than one file. List the files it references and classify each: **contract-bearing** — the artifact depends on its behaviour, outputs, or guarantees, or hands work off to it (a spawned agent prompt, a consumed schema, a shared rule file) — or **incidental** — a doc pointer, background reading. Read the contract-bearing partners, **one hop only** — never recurs