Brainstorming — Method Selection & Execution

What this skill does

Searches the 60-method brainstorming catalog (methods.csv) and applies the best-fit method to the user's problem. The agent acts as a facilitated brainstorming session conductor, not a passive method recommender.

Persona delegation

This skill delegates to @founder. The founder conducts the brainstorming session using the selected method. The skill file provides the method catalog and selection logic; @founder provides the Socratic depth, challenge stance, and structured reasoning.

When to use

• "Help me brainstorm ideas for..."
• "What brainstorming method should I use?"
• "Run a Six Thinking Hats session on my idea"
• Creative ideation before committing to a solution concept

Workflow

Step 1: Sharpen the topic (before any method)

Sharpening follows the shared contract .agents/references/step-0-sharpening.md (restate / fuzz ≤5 / emit / confirm — no manufactured ambiguity). Brainstorming-specific: the restated target names the output kind needed (ideas, risks, questions, alternatives); the confirmed statement drives method selection (passed as --context).

Step 2: Select method

Run:

node .agents/scripts/match_methods.js --source brainstorming --context "{sharpened topic}"

Return the best-matching methods with relevance scores (up to 5; fewer if the catalog match is narrow), each with a one-line why matched tied to the sharpened topic. Recommend the best one with reasoning. Let the user choose or accept the recommendation.

Step 3: Execute the method

Load the selected method from methods.csv. Apply it step-by-step:

• Announce in 1-2 lines why this method fits the sharpened topic and what it will test
• Explain the method's structure
• Specialist co-dispatch (mandatory on trigger): if the sharpened topic touches AI/ML (ML model, LLM prompt, RAG, ML eval, inference, AI agent behavior), security (secrets, auth, PII, trust boundary), or performance (latency, inference/compute-at-scale cost, load), spawn the matching critique voice as a native subagent for Step 3 — @ml-ai-engineer (Kai) for AI/ML, @security-engineer (Victor) for security, @performance-engineer (Felix) for performance. Bare words (model as in data-model, eval as in generic evaluation, cost as in budget, agent as in support-agent workflow) do NOT trigger — the ML qualifier must be present. The specialist challenges outputs mid-method (failure modes, cost, eval design), it does not conduct. Founder remains conductor. Log the co-dispatch ([CO-DISPATCH: @{seat} on {trigger}]); a triggered-but-absent seat needs [CO-DISPATCH-ABSENT: @{seat} — {one of: trigger-mentioned-but-out-of-scope | user-declined | solo-harness-no-subagents}] in the session output, counted per rolling 10 alongside routing overrides.
• Degraded dispatch (no subagents available) follows the shared contract .agents/references/degraded-mode-tiers.md — same tiers, --simulated exception rule, and telemetry as round-table and elicitation.
• Walk through each step interactively
• Record outputs (ideas, decisions, open questions)
• Report "What this surfaced" in 2-3 bullets tied to the sharpened topic
• Anti-sycophancy gate (SPC): if every generated idea is positive or agreeable — zero tension, zero risk, zero rejected assumptions — the session has converged prematurely. Before recording, apply a forced inversion pass (e.g., Reverse Brainstorming on the top idea: how would it fail?) and record the results alongside. Uniformly rosy output is an engine failure, not a success. Every session MUST emit exactly one marker line before the Step-4 write: [SPC: inversion-applied] (inversion ran, results recorded) | [SPC: rosy-absent — {quoted bullet id + rejected option | failure path | cost/risk tradeoff}] (the cited bullet must contain a rejected option, an explicit failure path, or a quantified tradeoff — "risk: adoption may be slow" does not qualify; otherwise inversion-applied is mandatory) | [SPC: skipped — {one of: user-explicit-override | fact-lookup-no-ideation | time-boxed-spike}] (closed taxonomy; logged to the routing record; skips above ~20% over a rolling 10 sessions trigger the same revisit as routing overrides). A Step-4 write without the marker fails step completion.
• Synthesize into actionable next steps

Step 4: Gate and write outputs

End the session with a Decision Gate on the strongest idea(s). The machine record — plain sentence, then the bracketed card — goes to artifacts/output/01-discovery/brainstorming-session.md. Format and depth: .agents/references/vespyr-dna.md §Decision Gate.

In plain wording the gate is: "Build this" ([GO], ready to stress-test), "Keep the why, replace the how" ([RESHAPE], promising but the mechanism needs rebuilding), or "Don't build this" ([NO-GO], dominated by alternatives). A brainstorm that ends without a gate is a list, not a session. The token is the machine companion, never the message.

Two surfaces. The chat reply carries the human sentence and the routing question — never [GO], [RESHAPE], [NO-GO], or any card field (.agents/references/vespyr-dna.md §Two surfaces). The token stays in the session file where the gates read it.

Step 4b: Route the gated idea (mandatory on GO/RESHAPE — do not menu)

The gate verdict locks FIRST — route from the locked verdict, never reshape the verdict to earn a panel. Count distinct domain tensions in the "What this surfaced" bullets, quoting bullet ids (e.g. tensions from B1/B3). Route from the document's verdict, then phrase the offer in plain words:

• [GO] + ≥2 tensions, or [RESHAPE] spanning ≥2 domains → roundtable (Discovery roster). Emit exactly: "Worth building, but {n} things are still unresolved (bullets {ids}: {a} vs {b}). Next: a roundtable to settle them. Reply 'grill-me instead' to go deep 1-on-1, or 'proceed' to start the roundtable."
• [GO] single-bet, or [RESHAPE] confined to one domain (state the domain + reason) → grill-me. Emit exactly: "Worth building, with one bet left to settle ({bet}). Next: grill-me to go deep on it. Reply 'roundtable instead' for a panel, or 'proceed' to start."
• [NO-GO] → archive + ledger. No build or follow-up routing (Zero-Blueprint rule blocks building, not talking — the user may still ask why it failed or re-propose with new evidence).
• Cost is part of the choice, state it once: grill-me costs one interviewer call; roundtable costs 2–5 isolated seats plus cross-examination rounds (5–10× tokens and attention). Single-bet/reversible → grill-me; multi-stakeholder/irreversible → pay for the panel.
• Pass artifacts/output/01-discovery/brainstorming-session.md path + the sharpened topic as the subject in either case. Grill-me = one mind going deep (proposer's conviction); roundtable = N independent minds colliding (what no single persona can think).
• Log every override as [ROUTING-OVERRIDE: default -> chosen -- reason] in the session document and the [BRAINSTORM] memory entry — append-only routing record, never in chat (owner @tech-lead counts markers per rolling 10 sessions alongside the router log). If overrides exceed ~20% over a rolling 10 sessions, the threshold rule gets revisited (recalibrate count vs kill default).

Write the session output — sharpened topic, method used, key ideas, gate verdict, open questions — to artifacts/output/01-discovery/brainstorming-session.md, then record to memory:

@memory-controller write active-decisions.md
### [BRAINSTORM] {method name} — {topic} [date: YYYY-MM-DD] [agent: @founder]
{synthesized outputs, key ideas, decisions, open questions}
Gate: {GO/RESHAPE/NO-GO} on {strongest idea}
**Status:** active

Output artifacts

• Session output written to artifacts/output/01-discovery/brainstorming-session.md
• Ideas and decisions written to artifacts/memory/active-decisions.md
• Brainstorming feeds into subsequent skills (/validate-idea, /shape-up, /design, /round-table and /grill-me via Step 4b routing)

Done when — checked mechanically

This skill has a real Decision Gate and a real routing rule, and until now nothing checked any of it. grep -c _check.js over this directory returned zero. The session artifact was written from a paragraph of instructions with no template, so the gate the skill defines could be absent, misspelled, or contradictory with the route it took, and nothing would notice.

node .agents/scripts/brainstorming_session_check.js check \
  --file artifacts/output/01-discovery/brainstorming-session.md --json

The SPC rule, and the trap it fell into first

Step 3 asks the agent to judge its own output: "if every generated idea is positive or agreeable — zero tension, zero risk, zero objection…". That is a self-assessment in the same field it fills in, which is the F-13 defect this repo has been repairing all session. So it is gated on the record.

The first version of that check passed a sycophancy session whose own note read "Zero tension. Zero risk. Zero objection. This is a great session." — because the note quoted the vocabulary the check counts. A document that declares the absence of the thing you are looking for is telling you it knows what you are looking for. Two guards now: the self-assessment region is excluded from the count, and a line asserting absence does not count as presence.

An honest session must be able to pass. Risk:, Trade-off:, Unknown:, open question, unvalidated, no data, failure mode, could fail all count — and the vocabulary is deliberately phrase-based, not an enumerated list, because an enumerated list rejects the honest phrasing nobody thought of and the cheapest response to a rejection is to change how you say it.

A session that found no tension may say so — if it says what it looked for. The first version of this check REQUIRED a tension, which forced a session that honestly found none to write a fabricated Risk: to pass. That is F-13 with the sign flipped: a gate whose cheapest passing answer is to invent evidence is worse than no gate, whichever direction it invents in. The fix is the one discovery_canvas_check.js already uses for quoted material — check the honesty of the declaration, not its direction:

The sycophant's answer to "is this sycophantic?" is not acceptable just because an honest answer is now allowed. The distinction is not which way the session leans — it is whether the claim can be checked. Two further traps found while testing it: a section heading named after the vocabulary (## Tension) counted as an instance of it, because the word is in the list; and a line quoting the vocabulary to assert its absence ("Zero tension. Zero risk.") counted as presence.

What this gate does not check

Whether the ideas are any good, whether the method was applied correctly, whether the verdict is right, or whether the session was worth having. Those need a human. It also cannot tell a session that honestly surfaced tension from one that performed it — only that the record does not claim effortless success.

Session template

# Brainstorming Session — {topic}

**Date:** YYYY-MM-DD
**Sharpened topic:** {Step 1's restated target — what is under discussion, and what output is needed}
**Method:** {one of the 60 method_name values in methods.csv}

## Key Ideas
1. {idea} — {why it could work}

## Decisions
- {what was decided, and on what basis}

## Open Questions
1. {unresolved question} — **Risk:** {what goes wrong if the assumption is false}
2. **Trade-off:** {what you give up}
3. **Unknown:** {what nobody has checked}

[SPC: inversion-applied]

**Verdict:** {plain sentence: build this / keep the why, replace the how / stop}

**Gate:** `[{GO|RESHAPE|NO-GO}]`
**Route:** {roundtable (≥2 tensions) | grill-me (single bet or one domain) | archive + ledger (NO-GO)}

Exit 0 conformant · 1 failures · 2 usage/IO error, fail-closed.

State machine integration

At start: Run node .agents/scripts/orchestrator_state.js status At end: Run node .agents/scripts/orchestrator_state.js complete --agent founder --artifact 01-discovery/brainstorming-session.md

Handoff

• Ideas ready for testing → load /validate-idea
• Semi-structured plan → load /shape-up
• Need market context → load /explore-idea
• Stress-test a GO/RESHAPE idea with a panel → Step 4b routes to /round-table (multi-tension) or /grill-me (single-bet)