Idea to Spec

Turn a rough idea into a short, evidence-backed spec made only of decisions the user made or confirmed. It has four parts: starting point, goals and non-goals, constraints, and proof.

Sequence

The phases are gates. Never skip one silently.

1. Homework. Follow references/homework.md before asking question 1.

2. Draft and critique. Spend cheap agent effort before expensive user attention.

   • Skip only when a wrong build costs a rerun or an unpick (throw it away, or undo some code). A migrate or can't-undo change (data already written, a public API, a sent message) always gets this step. Say which rung of the rework ladder it is (rerun, unpick, migrate, can't undo) and why, in one line.
   • Draft the smallest spec from homework facts alone, in the spec's four parts, with every assumption marked as one and every policy value marked (proposed). The draft decides nothing.
   • Critique: give the draft and the homework facts (each with source and date) to a fresh-context agent. Not a fork, not yourself, and without your reasoning. It hunts contradictions, unverified premises, silent decisions, missing actors, and arithmetic that breaks real limits, and returns its findings to you.
   • Fold each finding in as a homework re-check, a coverage-map area, or an interview question. Nothing from the draft reaches the spec without a decision from the user.

3. Interview. Follow references/interview.md: one decision per turn, 2 to 3 options, one recommendation.