Brainstorm
Diverge, then converge, then hand the winner to sw:increment. This never replaces
the spec — it decides which thing to spec.
Depth
| Depth | Options | Output |
|---|---|---|
quick (default for a narrow question) |
3 | a table + a pick, in the conversation |
standard |
4–5 | table + trade-offs + reports/brainstorm.md if an increment exists |
deep |
5–7, incl. one "do nothing" and one contrarian | written doc + rejected alternatives + risks |
Steps
- Frame. One sentence: the decision to be made, and what would make an answer good. Write down 3–5 criteria up front (cost, time-to-ship, blast radius, reversibility, who maintains it). Choosing criteria after seeing the options is how bias sneaks in.
- Diverge. Generate the options before judging any of them. Force variety:
- the obvious one (do the thing directly);
- the cheap one (what buys 80% for 20%);
- the buy-instead-of-build one;
- the do-nothing / defer one, with what it costs;
- one that inverts an assumption everyone is making. Name each option in 3–6 words.
- Converge. One table: option × criteria, with a one-line "kills it if" per option. Then a short paragraph per surviving option: how it works, what it costs, what breaks.
- Pick. State the recommendation, the runner-up, and the one fact that would flip the decision. If the honest answer is "we need data first", say that and name the experiment.
- Hand off.
sw:increment "<the picked option>". Paste the rejected alternatives into the new spec's Approach section — that is where they earn their keep.
Rules
- Never grade an option while still generating options.
- Every option gets the same amount of scrutiny; do not build a straw man to justify the favourite.
- Tables and short paragraphs, not essays.
- Persist the doc only when there is an increment to hold it
(
.specweave/increments/<id>/reports/brainstorm.md); otherwise keep it in the conversation. - If the user already knows what they want, skip to
sw:incrementand say why.
Multi-perspective variant
For a genuinely contested decision, run three lenses over the same option set — advocate
(strongest case for), critic (how each fails in production), pragmatist (what ships this
week) — and merge. In Claude Code these can be parallel subagents (sw:team in brainstorm
mode); in other tools, do the three passes in sequence and label them.