bfw — Brainstorming Framework

A facilitated ideation skill for Claude Code (and compatible AI coding assistants). Drives structured brainstorming sessions from raw ideas through crystallised decisions, using classical techniques (Six Hats, SCAMPER, MoSCoW, Impact/Effort, First Principles, and more) with adaptive sequencing.

Why

Most AI brainstorming sessions drift: you start with a question, dump some ideas, the AI agrees with everything, and you end with a longer list of unordered thoughts. bfw treats the AI as a facilitator, not a participant — it drives a structured process, keeps energy high, detects convergence or stuck signals, and adapts the plan mid-session.

The key design decisions:

• Announce + adapt without asking permission. A good facilitator changes course when the room shifts. Constantly asking "is that OK?" breaks trust and flow.
• Problem shape detection. The same generic "brainstorm" template doesn't fit a blank-slate question and a 30-item backlog. bfw classifies the input into 5 shapes and picks a recipe accordingly.
• Meta-analysis is first-class. Every session ends with "what worked, what didn't, what to try next time" — feeding back into the technique library for continuous improvement.
• Encouraging but honest. The facilitator celebrates good thinking, but emits unprompted criticism when it spots blind spots, missing constraints, or unchallenged assumptions.

What's in the box

bfw/
├── .claude-plugin/
│   ├── plugin.json              # Claude Code plugin manifest
│   └── marketplace.json         # Single-plugin marketplace catalog
├── skills/
│   └── brainstorm/
│       ├── SKILL.md             # The main skill prompt — the brain
│       └── techniques.md        # Technique library (20 techniques, 5 phases)
├── README.md                    # This file
├── CHANGELOG.md
├── CLAUDE.md                    # Orientation for Claude Code sessions
├── LICENSE                      # MIT
├── justfile                     # Release automation
└── examples/                    # Real session outputs
    └── site-ideas-prioritization.md

Installation

bfw ships on two parallel install channels from the same repo, so you can use it whether your assistant speaks Claude Code plugins or the cross-host skills CLI.

Claude Code (recommended)

/plugin marketplace add bastien-gallay/bfw
/plugin install bfw@bfw

This installs the bfw plugin from its own marketplace and exposes the brainstorm skill. Claude can invoke it automatically from the description, or you can call it explicitly via /bfw:brainstorm. Claude Code auto-updates the plugin at session start.

Everywhere else (45+ hosts via skills)

npx skills add bastien-gallay/bfw

Works in Cursor, Codex, Cline, Continue, Copilot, Windsurf, Warp, Gemini CLI, Roo Code, and dozens more. Run npx skills update periodically to pull the latest version — unlike the Claude Code plugin path, updates here are manual.

Claude Desktop / claude.ai (web)

Download the latest bfw-brainstorm-vX.Y.Z.skill from the releases page.

• On Desktop: double-click the file. Claude Desktop opens an "Add to your library?" dialog previewing the skill — click Add to library and you're done.
• On the web: open Settings → Capabilities → Skills, click + Create skill, and upload the .skill file.

.skill is a zip archive with an OS-registered extension; Desktop installs it on double-click without a Settings trip. Updates are manual (download the new release, re-install). Available on Pro, Max, Team, and Enterprise plans with code execution enabled.

Manual / other hosts

bfw is just two markdown files under skills/brainstorm/. Any assistant that can load a system prompt or skill file can use it — point it at skills/brainstorm/SKILL.md (and keep techniques.md next to it).

Usage

/brainstorm TOPIC="Should we rewrite the legacy API?" DURATION=15

/brainstorm \
  TOPIC="Q3 product priorities" \
  IDEAS="AI search, mobile app, API v2, dashboard redesign" \
  DURATION=25

What happens:

1. Intake — extends your ideas, adds 2-4 AI suggestions, asks for constraints
2. Shape — detects whether you have too few / too many ideas, a binary choice, or decision under constraints
3. Plan — announces a technique sequence (Six Hats, MoSCoW, Impact/Effort, etc.) with estimated times
4. Execute — runs each technique, presents results, collects feedback (closed + open questions)
5. Reflect — after each step, adapts the plan if the session signals convergence / stuck / new dimension
6. Crystallise — selects final ideas with rationale, generates action items
7. Meta-analysis — records what worked for future improvement

Sessions are persisted as markdown under the configured output directory (default: brainstorm/YYYYMMDD-<slug>.md).

Problem shapes and default recipes

Input shape	Trigger	Recipe
Blank slate	0-2 ideas	Starbursting → Free Association → Affinity Group → NUF → Action Items
Few ideas	3-7 ideas	SCAMPER → Impact/Effort → Action Items
Many ideas	8+ ideas	Affinity Group → MoSCoW → Action Items
Large backlog	15+ ideas	Affinity → Dot Vote (cluster) → Impact/Effort (filtered) → MoSCoW
Binary choice	Exactly 2 ideas	PCI → Devil's Advocate → Binary Comparison
Decision under constraints	Constraints mentioned	Constraint Mapping → Impact/Effort → MoSCoW → Action Items

Recipes can always be overridden with TECHNIQUES="six-hats,moscow". Six Hats is opt-in only — request it explicitly via TECHNIQUES or ask for stress-testing at intake.

Technique library

20 techniques across 5 phases. See skills/brainstorm/techniques.md for the full reference.

Diverge — Starbursting, SCAMPER, Reverse Brainstorm, What-Would-X-Do, Random Stimulus, Free Association Analyze — Six Thinking Hats, First Principles, 5 Whys, SWOT, PCI Converge — MoSCoW, Impact/Effort Matrix, NUF Test, Decision Matrix, Binary Comparison, Dot Voting Crystallize — Action Items, Elevator Pitch Adaptive — Affinity Grouping, Devil's Advocate, Constraint Mapping

Customization

Add your own techniques by editing techniques.md. Each entry is a markdown section with:

• Purpose — what it's for
• When — when to use it
• How — step-by-step instructions for the AI facilitator
• Duration — typical time budget
• Output shape — what the result looks like

The skill reads techniques.md at session start, so new techniques are available immediately.

Session output example

See examples/site-ideas-prioritization.md for a real session that prioritised 30 ideas for a documentation site into a Must/Should/Could/Won't plan, with meta-analysis.

Design philosophy

Four principles drove the design:

1. Structure beats freeform for AI brainstorming. LLMs are too agreeable on their own. Imposing a process forces real analysis.
2. Adaptation beats rigid plans. The best facilitators change course when the room shifts. The framework detects and announces changes transparently.
3. Double Diamond over single-pass. Every session alternates divergence (opening) and convergence (closing), mirroring the classic design-thinking pattern.
4. Meta-analysis is the moat. Recording what worked turns every session into a data point for improving the next one.

License

MIT — see LICENSE.

Credits

Created by Bastien Gallay. Contributions welcome.