brainstorm-internal

Some choices are not the user's. Which of two implementations satisfies a contract better, how to sequence four increments, what an ADR should record as rejected — the user has nothing to add, and asking costs a turn and gets "you decide". This is the divergence technique for those.

There is no user here. No questions, no pacing, no cap, no confirmation to wait for. The nine-question machinery in docs/graph/skills/brainstorm-socratic.md is not a stricter version of this skill — it is the other mode, and it does not apply when there is nobody to ask. The exit is a written options set, not an agreement.

The failure this exists to prevent

A session brainstorming against itself generates one real option and two strawmen. It has already, quietly, decided; what it produces is the decision plus two alternatives shaped to lose. The output looks exactly like genuine divergence — three options, a comparison, a pick — and an ADR built on it records rejected alternatives nobody ever considered.

Every rule below exists because of that one failure.

The method

1. Write the decision as a question with at least two answers you could defend. If you cannot defend the second answer, you have not found an option yet; you have found an objection to one. Go back.

2. Generate against evidence already in hand — the ledgers, the wiki pages, the specs, the source. This mode does not go to the internet (agent.research-scout does) and does not ask the user (skill.brainstorm-socratic does). If a genuine option needs evidence you do not have, that is the finding: name it and say which of the two it would settle.

3. For each option, state what would have to be true for it to be right. Not its advantages — its preconditions. This is the step that kills strawmen, because an option whose preconditions you cannot write down is one you invented to lose.

4. For each option, name the single fact that would kill it. If an option has no kill condition it is not a real option; if two options share a kill condition they are one option described twice.

5. Say which preconditions are already known true, known false, or unchecked. Unchecked is the useful category and the one that gets collapsed: an option rejected on an unchecked precondition is rejected on a guess, and must be recorded as such.

6. Pick, and record why the others lost — against their kill conditions, not against their vibes. The losers go into grill.md §7 or the ADR's rejected alternatives verbatim; that is what makes this mode's output worth producing rather than reasoning silently.

Honesty checks before you exit

• Could a competent reader of the losing options tell they were seriously considered? If the answer depends on your say-so, they were not.
• Did any option change shape while you worked? Real divergence moves; a set that arrives finished was a decision wearing three hats.
• Is any option rejected on a precondition you never checked? Say so in the record, or check it.
• Did you produce exactly three options because three is the habit? Two defensible options beat three where one is filler.

Exit conditions

• Each option carries its preconditions and its one kill condition.
• Every precondition is marked known-true, known-false, or unchecked.
• The pick names which kill condition retired each loser.
• The options set is written where the next step reads it — grill.md §7, an ADR's rejected alternatives, or the plan record — not left in the session.

No user confirmation is required to exit. If it turns out one is — the decision was the user's after all — you were in the wrong mode: brainstorm.mode-selection in docs/graph/protocols/brainstorm.md is the test, and switching is cheap while the options set is fresh.