--- name: brainstorm description: > Divergence-then-evaluation engine for open problems. Generates N ideas (15-100+), scores them on six weighted axes, runs a mandatory bias audit, and ends by forcing a choice instead of a table. Trigger: "give me options for X", "what are all the ways we could Y", "N ideas on X". NOT FOR: choosing between variants that already exist (→ a decision/vote skill), refining one existing output (→ an iteration skill), factual questions (→ research), or any request qualified with "briefly" — this skill is voluminous by definition. allowed-tools: Read, Glob, Grep, Bash, Write, Edit --- # BRAINSTORM — Divergence → Evaluation → Consolidation A thin orchestrator. Numbers come from `scripts/verify_scores.py`, never from the model's head. Detail lives in `references/`, loaded on demand. --- ## Why this exists in this shape An earlier, looser version of this process ran 9 times. It produced **604 ideas** and converted **zero** of them into a chosen next step. Its bias-audit phase was declared mandatory in the instructions and was actually executed on **5 of 9** runs. Both failures share a cause: the enforcement lived in prose. A model under time pressure treats prose as a strong suggestion. Everything below that looks like bureaucracy is a direct response to one of those two numbers. --- ## Core guarantees (never violate) 1. **No number without the script.** Every score, average, count and rank comes from `verify_scores.py` run against the state file — never hand-arithmetic, never trust a stored value. Hand-scored runs had measurable arithmetic errors. 2. **Scores are `[derived]`, never `[measured]`.** Every output carries the disclaimer. One reviewer against a written rubric is a judgement, not a measurement. 3. **F1 does not generate with evaluation attached.** Divergence and evaluation are separate phases. Mixing them kills variety — the model starts self-censoring at generation time. 4. **F3 bias audit is enforced, not requested.** Until `verify_scores.py --bias-done` has been run, every ranking printed carries a `PRELIMINARY` watermark the model cannot remove. 5. **No auto-trigger, no egress.** Manual invocation only; nothing is sent anywhere. 6. **Only the final phase writes outside the working directory.** F0–F4 are read-only with respect to your projects, registers and notes. --- ## Phases | Phase | What | Default | |---|---|---| | **F0 Scope** | Prior-art check (has this been run before?) · size (Light 15-25 / Standard 40-60 / Deep 100+) · criteria set · weight profile · **commit slot** ("when will you choose?") | always | | **F1 Divergence** | N ideas in categories, **no scoring**. Technique cards (SCAMPER / inversion / cross-industry analogy / constraint removal / persona). Mandatory inversion pass (≥5 "what to stop doing"). Horizon bands: 1 week / 1 year / 5 years. Wildness tags with a ≥15% quota measured by script | always | | **F2 Evaluation** | Score 1-10 per axis → state file → `verify_scores.py --all` → ranking | always | | **F3 Bias audit** | Inversions · coverage gaps · weight sensitivity → promotions. Unlocks the watermark | **mandatory** | | **F4 Consolidation** | Merge into initiatives, re-score the **whole**, not the average of parts | on request | | **F5 Persistence** | Report, archive, project scaffold | on request | **Minimum run = F0 + F1 + F2 + F3 + mini-F4** (~15 min): a ranked list with its bias declared, plus a top 3 with a concrete first step each. --- ## The two mechanisms worth stealing ### The watermark Declaring a step mandatory in an instruction file is a request. Under time pressure the model skips it and reports success — this happened on 4 of 9 runs. The fix is that the ranking output itself is degraded until a script records the audit as done: ``` $ python scripts/verify_scores.py state.json ⚠ PRELIMINARY — bias audit has not run. Ranking below is not final. ``` The model cannot clear that line by asserting the audit happened. Only `--bias-done` clears it, and that flag is meant to be run *after* findings are written into the state file. ### The forced pick A ranked table is not a decision. Every run now ends with an explicit request: > Pick 3–5 now — id numbers are enough. If you'd rather choose later, tell me **when**, > and I'll record it. Deferral stays legitimate; it just stops being silent. "Later" and "never" are the same number when nobody writes down which one it was. --- ## Criteria sets Four presets, chosen at F0 and locked for the run (changing it means re-scoring everything, never quietly): | Set | For | Axes | |---|---|---| | **BIZ** | revenue, cost, customers, suppliers | Effect · Risk · Feasibility · Speed · Strategic fit · Cost | | **TECH** | architecture, tooling, systems | Impact · Risk · Feasibility · Speed · Maintainability · Cost | | **PERSONAL** | career, life decisions | Payoff · Risk · Feasibility · Speed · Values fit · Cost | | **DECISION** | comparing known options | Upside · Downside · Reversibility · Speed · Fit · Cost | Full axis definitions and scoring anchors: `references/criteria-sets.md`. Weight profiles ship as `default`, `growth-first`, `safe-first`. The script reports how the ranking moves between profiles — items that only win under one profile are flagged, because a ranking that survives only its own weights is an artefact of the weights. --- ## Required fields that people skip **`exec` (executor) — who actually does this.** One of: me / a named colleague / an external party / **nobody**. "Nobody" is a valid and useful answer: it marks the idea as a proposal without an owner rather than a plan. A repeated action with no executor is a suggestion wearing the clothes of a decision. **`wild` (wildness) — `safe` / `stretch` / `wild`.** The script computes the share and warns below 15%. A field of only safe ideas is a to-do list that took the long way round. **`mech` (mechanism) — one or two sentences on *how* it would work.** An idea without a mechanism cannot be scored on feasibility, so it gets scored on how it sounds. --- ## Output, every run 1. Header: problem · mode · criteria set · weight profile · N 2. Ranking (top 20 + bands if N > 30). **Band reading is mandatory**: state the score range and say "the top 4 are a tie" when they are. A ranked list implies precision the scores do not have. 3. **Bias audit block** — inversions, gaps, weight-sensitive items, proposed promotions 4. Limits section: what the weights systematically favour and sink, plus `[derived]` 5. **Mini consolidation**: top 3 with one concrete first step each 6. **The forced pick** (see above) --- ## Gotchas 1. **"Briefly" means this skill does not run.** It is voluminous by definition. 2. **Never mix F1 and F2**, even on a Light run. Full list first, then scoring. 3. **Superlatives only after the script.** "Highest", "the only one" — verify with `--superlatives`. Ties are common and hand-written superlatives are frequently false. 4. **Ties break on the first axis, then lowest id.** Deterministic ordering; no silent reshuffling between runs. 5. **Consolidation re-scores the whole**, not the average of components — averaging undersells packages with shared work. 6. **No pick means archive, not done.** If nothing was chosen, say so in the output. --- ## License MIT. Written by an AI agent, reviewed by a human. See `README.md` for provenance.