# bfw โ 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** โ 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" โ 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
```text
bfw/
โโโ .claude-plugin/
โ โโโ plugin.json # Claude Code plugin manifest
โ โโโ marketplace.json # Single-plugin marketplace catalog
โโโ skills/
โ โโโ brainstorm/
โ โโโ SKILL.md # The main skill prompt โ the brain
โ โโโ techniques.md # Technique library (20 techniques, 5 phases)
โโโ README.md # This file
โโโ CHANGELOG.md
โโโ CLAUDE.md # Orientation for Claude Code sessions
โโโ LICENSE # MIT
โโโ justfile # Release automation
โโโ examples/ # Real session outputs
โโโ 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)
```bash
/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`)
```bash
npx skills add bastien-gallay/bfw
```
Works in Cursor, Codex, Cline, Continue, Copilot, Windsurf, Warp,
Gemini CLI, Roo Code, and [dozens more](https://github.com/vercel-labs/skills).
Run `npx skills update` periodically to pull the latest version โ
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](https://github.com/bastien-gallay/bfw/releases/latest).
- **On Desktop:** double-click the file. Claude Desktop opens an
**"Add to your library?"** dialog previewing the skill โ click
**Add to library** and you're done.
- **On the web:** open **Settings โ Capabilities โ 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 โ
point it at `skills/brainstorm/SKILL.md` (and keep `techniques.md`
next to it).
## Usage
```bash
/brainstorm TOPIC="Should we rewrite the legacy API?" DURATION=15
```
```bash
/brainstorm \
TOPIC="Q3 product priorities" \
IDEAS="AI search, mobile app, API v2, dashboard redesign" \
DURATION=25
```
**What happens:**
1. **Intake** โ extends your ideas, adds 2-4 AI suggestions, asks for
constraints
2. **Shape** โ detects whether you have too few / too many ideas, a
binary choice, or decision under constraints
3. **Plan** โ announces a technique sequence (Six Hats, MoSCoW,
Impact/Effort, etc.) with estimated times
4. **Execute** โ runs each technique, presents results, collects
feedback (closed + open questions)
5. **Reflect** โ after each step, adapts the plan if the session
signals convergence / stuck / new dimension
6. **Crystallise** โ selects final ideas with rationale, generates
action items
7. **Meta-analysis** โ 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 โ Free Association โ Affinity Group โ NUF โ Action Items |
| **Few ideas** | 3-7 ideas | SCAMPER โ Impact/Effort โ Action Items |
| **Many ideas** | 8+ ideas | Affinity Group โ MoSCoW โ Action Items |
| **Large backlog** | 15+ ideas | Affinity โ Dot Vote (cluster) โ Impact/Effort (filtered) โ MoSCoW |
| **Binary choice** | Exactly 2 ideas | PCI โ Devil's Advocate โ Binary Comparison |
| **Decision under constraints** | Constraints mentioned | Constraint Mapping โ Impact/Effort โ MoSCoW โ Action Items |
Recipes can always be overridden with `TECHNIQUES="six-hats,moscow"`.
Six Hats is opt-in only โ request it explicitly via `TECHNIQUES` or
ask for stress-testing at intake.
## Technique library
20 techniques across 5 phases. See
[skills/brainstorm/techniques.md](skills/brainstorm/techniques.md) for
the full reference.
**Diverge** โ Starbursting, SCAMPER, Reverse Brainstorm, What-Would-X-Do,
Random Stimulus, Free Association
**Analyze** โ Six Thinking Hats, First Principles, 5 Whys, SWOT, PCI
**Converge** โ MoSCoW, Impact/Effort Matrix, NUF Test, Decision Matrix,
Binary Comparison, Dot Voting
**Crystallize** โ Action Items, Elevator Pitch
**Adaptive** โ Affinity Grouping, Devil's Advocate, Constraint Mapping
## Customization
Add your own techniques by editing `techniques.md`. Each entry is a
markdown section with:
- **Purpose** โ what it's for
- **When** โ when to use it
- **How** โ step-by-step instructions for the AI facilitator
- **Duration** โ typical time budget
- **Output shape** โ 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](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 โ see [LICENSE](LICENSE).
## Credits
Created by Bastien Gallay. Contributions welcome.