The brainstorm skill forces you past the obvious ideas by requiring every new idea to be structurally distinct from all previous ones. It's like having a creative partner who never lets you repeat yourself.
You need to design a notification system for a developer platform. The obvious approaches (email, Slack webhook, in-app toast) come to mind immediately — but are they the best?
/brainstorm 15 How should we notify developers about build failures, deployment status, and security alerts in our platform?
The 15 sets the target to 15 ideas (default is 30). Claude will:
- Initialize the session with a unique ID
- Research existing notification patterns (Phase 2)
- Catalog prior art — email, Slack, SMS, push notifications, etc. become "banned territory"
- Begin divergence — now you generate 15 ideas that are structurally different from all prior art AND from each other
Claude searches for existing notification approaches and catalogs them:
[x] Email notifications (traditional async delivery)
[x] Slack/Teams webhooks (channel-based real-time)
[x] Push notifications (mobile/desktop OS-level)
[x] In-app notification center (feed/inbox pattern)
[x] SMS/pager (high-urgency escalation)
[x] RSS/Atom feeds (pull-based subscription)
6 prior art entries total — all banned during brainstorming.
Now you (and Claude, as a brainstorming partner) generate ideas. Each one goes through a uniqueness gate:
> Idea: "Ambient dashboard that changes color based on system health"
#01 Ambient dashboard that changes color based on system health
[████████░░░░░░░░░░░░] 1/15 14 to go
If you propose something too similar to an existing idea:
> Idea: "Dashboard with status indicators"
Too close to #01 "Ambient dashboard that changes color...".
The shape needs to change, not just the surface.
Consider: What if the solution had to work without any visual UI at all?
The system provides varied nudges and constraint-shifting questions to push you into new territory.
After hitting 15 ideas, rate each on Feasibility, Novelty, and Impact (1-5):
/brainstorm rate <session> 7 4 5 3
#07 Git-blame-aware alerts that only notify the person whose code broke
F=4 N=5 I=3 total=12/15
Generate the final report to see rankings, hidden gems, and quick wins.
The core innovation. "Shape" means the structural approach — not surface details. Two ideas that would produce the same flowchart are the same shape, even if they use different technologies.
The system enforces this with multiple layers:
- Lexical: word and bigram overlap detection
- Semantic: embedding-based similarity (requires
sentence-transformers) - Structural: FTS5 keyword boosting for borderline cases
- Code: if you prototype ideas in sandboxes, code structure is also compared
For each idea, you can create a sandbox to build a minimal prototype:
sandbox.sh create <session> <idea-number>Then run check-code to verify the prototype is structurally different from other prototypes. This catches ideas that sound different in English but produce identical code.
Prior art isn't just documentation — it actively shapes the brainstorm. By mapping existing approaches and banning them, you force yourself into genuinely new territory. The research phase isn't busywork; it's building the boundaries that make divergence productive.
- Start with a high target (20-30) if you're exploring a wide problem. You'll be surprised how creative you get once the obvious ideas are exhausted.
- Let Claude propose ideas too. It's a brainstorming partner, not a scribe.
- Don't filter during divergence. Bad ideas at position 8 often inspire good ideas at position 12.
- Use the constraint questions. When the system asks "What if you only had 1 GB of memory?", take it seriously — constraints unlock new solution shapes.