plan-with-teeth

The rule

A plan is a stack of claims about code, and every claim is either read from the source this session or it is a guess. Plans built on guesses don't fail at the whiteboard; they fail at step 3 of the implementation, where rework is expensive. Read first, label what you didn't read, and attack your own draft before anyone else sees it.

This sharpens superpowers' writing-plans: that skill gives a plan its shape and process, this one keeps its content honest.

Fires when

Writing an implementation plan or design note, entering plan mode, answering "how would we build this?", estimating the size of a change, reviewing someone else's plan.

How to apply

1. Read every file the plan touches before writing a step about it. A claim about the codebase carries a file:line reference. Not read this session means not known.
2. Open the plan with an assumptions ledger. Mark each entry [VERIFIED file:line] or [HYPOTHESIS]. A load-bearing hypothesis gets resolved by reading the code or asking; it never gets built on silently.
3. Bring two or three genuinely different approaches, with cost, risk, and blast radius. Say why the losers lose. A one-approach plan is a draft.
4. End every step with how you'll know it worked: a command, a test, an observable behavior. A step with no check is a hope.
5. After drafting, attack the plan: which edge case kills it, which file did you not open, what does rollback look like, what would a reviewer bounce. Fix what you find, then show it.

Anything still unverified in the final plan stays labeled a hypothesis. A wrong plan costs more than a slow one, so think as long as it takes.

Worked example

Task: "plan rate limiting for our public API." The plan from memory says: add middleware, configure limits, write tests. Three steps, one approach, nothing checked. The plan with teeth starts by reading the request path and finds a response cache sitting above the middleware chain (gateway/cache.ts:88), so naive middleware would count cache hits against the limit. Its ledger holds "[HYPOTHESIS] every endpoint goes through the main router" — checked, and false: the websocket upgrade path skips it. It weighs an in-process token bucket against gateway-level limiting and picks the gateway, because the service runs six replicas with no shared state. Each step ends with a check ("hit the endpoint 30 times, read the 429 body and the Retry-After header"). The cache and the websocket path are exactly the two surprises the from-memory plan would have met mid-implementation.