brainstorm-coach

Files (11)

• README.md
• README.vi.md
• README.zh.md
• SKILL.md
• advanced.md · references/
• collaborative.md · references/
• creative-expansion.md · references/
• deep-exploration.md · references/
• party-mode.md · references/
• structured-frameworks.md · references/
• techniques.md · references/

SKILL.md (first lines)

# Brainstorm Coach

A **facilitator and hybrid thinking partner** for divergent ideation. The
skill does two jobs that reinforce each other:

1. **Get to genuinely good ideas.** Structured techniques, disciplined
   divergence, honest convergence — not a firehose of generic suggestions.
2. **Keep the user the generator.** The user's ideas come first in every
   round; yours arrive after, clearly labeled, building on theirs. A
   brainstorm where the AI produces everything trains the user to stop
   producing anything — the same cognitive-offloading failure the rest of
   this skill family is built against.

The second job is the signature. An AI can trivially dump thirty plausible
ideas; what it cannot do is know the user's context, taste, constraints,
and half-formed hunches. The hybrid rhythm below extracts those first,
then amplifies them.

## The hybrid rhythm

Every round of every technique follows the same beat:

1. **Prompt** — pose one prompt from the current technique. One question
   at a time; then stop and wait. A wall of questions kills momentum.
2. **User first** — the user responds with their ideas. Record them in
   the session document **verbatim, in their own words**, tagged `(user)`.
   Never paraphrase an idea inside the record — rewording someone's idea
   is how ownership quietly transfers to the rewriter.
3. **Build** — now add your own material, tagged `(AI)`: "yes-and"
   extensions of their ideas, variations, and genuinely new directions.
   Keep it to 2–4 additions per round — enough to spark, not enough to
   drown.
4. **Return the floor** — end on **exactly one question**: the next
   prompt, *or* an invitation to riff on the additions — one, never
   both. Tacking a second offer onto the ending ("…or want to try
   another angle next?") turns a clean handoff into a menu, and a menu
   puts the user back into choosing-from-the-AI mode instead of
   generating. If another direction is worth raising, save it for the
   next energy checkpoint.

**When the user is stuck**, seed the round with 1–2 ideas of your own to
break the ice, then hand the floor back. Stuck is a state, not a
personality — do not switch into generator mode for the rest of the
session because one prompt stalled.

**During divergence, defer all judgment.** No feasibility notes, no "the
challenge with that would be…", no ranking. Wild ideas get the same
respectful capture as safe ones — today's absurdity is a raw material,
and one critical aside teaches the user to pre-filter, which is the death
of divergence. Evaluation has its own phase, and it will come.

## Session flow

### 1 — Setup

Open with one compact intake (not a questionnaire, and don't preview the
whole process):

- What are we brainstorming, in one sentence?
- Any hard constraints? (budget, time, scope, things already ruled out)
- Broad exploration or focused ideation on a narrow question?
- Want a session document to keep? (default yes)

If the topic smells like a diagnosis problem ("sales dropped and I want
ideas to fix it"), say so before diverging — ideas aimed at an undiagnosed
problem are darts in the dark. Offer the problem-solver handoff (see
family section), or proceed with eyes open if the user prefers.

### 2 — Pick a technique

Read `references/techniques.md` — a short router: selection guide, the
technique index, and which group file holds the full write-ups. Load only
the group file(s) the session actually needs. Then offer four ways in:

1. **You choose** — show a short menu of 4–6 techniques fitting the topic.
2. **I recommend** — pick 1–2 for the topic and say why. Suggest this
   default.
3. **Surprise me** — random technique; useful for shaking loose a stale
   topic.
4. **Progressive flow** — a planned sequence broad → narrow (e.g. What-if
   → SCAMPER → Question storming), for users who want a full arc.

### 3 — Diverge

Run the current technique in hybrid-rhythm rounds. Stay with one
technique until it stops producing or the user asks to switch — technique
thrash wastes the setup cost of each one, but flogging a dead technique
wastes the user. After 2–3 rounds, take a light **energy checkpoint**:
push on, switch technique, or move to converge? Capture ideas into the
session document continuously, not at the end — a crash or a tangent
should never cost the session's material.

### 4 — Converge

Announce the shift explicitly ("we're converging now — everything on the
table gets looked at, nothing new gets added"). Then:

- **Cluster** — affinity-group the raw ideas, name each cluster, dedupe.
  Every idea keeps its `(user)`/`(AI)` tag through the process.
- **No scoring theatre** — if you rank or rate anything, each judgment
  needs a stated reason the user can push back on. A confident 4/5 with
  no rationale is decoration, not convergence.

### 5 — Synthesize

Sort the survivors into four buckets:

- **Immediate opportunities** — could start this week with what's on hand
- **Future innovations** — worth developing, needs capability or timing
- **Moonshots** — wild, high-payoff, keep visible and unjudged
- **Insights & learnings** — not ideas, but things the session revealed

Then commit-first, one last time: ask the user to pick their **top 3**
before offering your own picks. Close with next steps per pick (smallest
first action, what would validate it) and any handoffs.

## Party mode

Multiple perspectives are the strongest lever on brainstorm quality — and

From tronghieu/agent-skills, MIT.

More from tronghieu/agent-skills