Agent skill
brainstorm
Diamond Model brainstorming — diverge ideas without judgment, then converge with quantified evaluation. Wraps deep-interview for requirements clarification. Triggers on "brainstorm", "브레인스토밍", "아이디어", "ideate", "what if", "explore options".
Install this agent skill to your Project
npx add-skill https://github.com/devseunggwan/my-skills/tree/main/skills/brainstorm
SKILL.md
Brainstorm
Overview
Brainstorming expands possibilities before narrowing them down. Unlike planning (HOW) or deep-interview (WHAT exactly), brainstorming explores the WHAT IF space.
Core principle: Diverge fully before converging. Never judge during idea generation.
Delegates to: OMC architect agents (multi-lens ideation), analyst agent (clustering), critic agent (evaluation)
The Iron Law
NO JUDGMENT DURING DIVERGENT PHASE. NO SELECTION WITHOUT EVALUATION.
If you haven't completed Phase 1 (Diverge), you cannot evaluate or select ideas. If you haven't completed Phase 3 (Evaluate), you cannot recommend an idea.
When to Use
- Direction is unclear — don't know what approach to take
- Exploring possibilities for a new feature, architecture, or tool
- Stuck on a problem and need fresh perspectives
- Multiple valid approaches exist and need structured comparison
- User says "brainstorm", "아이디어", "what if", "explore options"
Use this ESPECIALLY when:
- "Just go with the obvious approach" feels tempting
- Only one option seems to exist (usually means insufficient exploration)
- The problem domain is unfamiliar
The Diamond Model
Phase 1: DIVERGE Phase 2-3: CONVERGE
╱ ideas expand ╲ ╱ evaluate ╲
╱ no judgment ╲ ╱ rank ╲
╱ quantity first ╲ ╱ select ╲
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
open wide narrow to best
Process
Phase 0: Context Gathering (Automatic)
Before any ideation, gather relevant context silently:
- Run
exploreagent (haiku): scan codebase for related patterns, constraints, prior art - Identify technology stack, existing conventions, and hard constraints
- If user provided a topic, search for similar solved problems in the codebase
Rule: Gather codebase facts BEFORE generating ideas. Ideas grounded in reality are better than ideas from vacuum.
Phase 1: Diverge — "Yes, And"
Generate as many ideas as possible. NO judgment. NO filtering. NO evaluation.
Multi-Lens Technique (run 3-5 architect agents in parallel):
| Lens | Perspective | Guiding Question |
|---|---|---|
| Analyst | Data/logic-driven | "What approaches are logically possible given our constraints?" |
| Contrarian | Opposite viewpoint | "What if we did the exact opposite? What if we intentionally made it worse?" |
| Analogist | Cross-domain | "How do other domains solve similar problems?" |
| Minimalist | Remove complexity | "What if there were no constraints? What's the simplest possible version?" |
| User | End-user perspective | "What does the actual user want? What would delight them?" |
Each lens generates 3-5 ideas → 15-25 candidates total.
Rules during Phase 1:
- Every idea is valid (no "that won't work")
- Build on others' ideas ("yes, and...")
- Wild ideas welcome (they often seed practical solutions)
- Quantity over quality at this stage
Phase 2: Cluster & Enrich
- Group similar ideas into clusters (3-7 clusters typical)
- Name each cluster with a clear theme
- Identify combinable ideas across clusters (SCAMPER: Combine)
- Fill gaps: "What perspective is missing?"
- Synthesize hybrid ideas from promising combinations
Phase 3: Evaluate & Rank
Score each candidate on a 4-axis matrix:
| Axis | Weight | Description |
|---|---|---|
| Impact | 35% | How much value does it create or how well does it solve the problem? |
| Feasibility | 30% | Can we build this with current tech, resources, and time? |
| Novelty | 20% | How different is this from existing approaches? |
| Risk | 15% | What's the downside if it fails? (inverse: lower risk = higher score) |
Scoring: 1-5 per axis → weighted average → rank.
Delegate evaluation to OMC critic agent (opus) for rigorous, unbiased scoring.
Phase 4: Output
Present results in structured format:
## Brainstorm Results: {topic}
### Context
{codebase findings, constraints identified in Phase 0}
### Top Ideas (ranked)
1. **[Idea Name]** — Score: X.X/5
- What: one-sentence description
- Why it scores high: key strengths
- Sketch: rough implementation approach
- Risk: main concern
2. ...
### Honorable Mentions
- Ideas that scored well but narrowly missed top ranks
### Discarded (with reasons)
- Ideas that were evaluated and rejected, with brief rationale
### Recommended Next Step
Phase 5: Execution Bridge
After presenting results, offer next steps:
- Deep Interview → Clarify requirements for the top idea
- Plan → Create implementation plan for a selected idea
- Refine → Run another brainstorm round with narrower scope
- Compare → Deep-dive trade-off analysis of top 2-3 ideas
Modes
| Mode | Flag | Behavior |
|---|---|---|
| Interactive (default) | none | User participates in each phase, adds their own ideas |
| Solo | --solo |
Agent runs all phases autonomously, presents final results |
| Workshop | --workshop |
5 parallel architect agents for maximum idea diversity |
OMC Agent Delegation
| Phase | Agent | Model | Role |
|---|---|---|---|
| Phase 0 | explore |
haiku | Codebase context gathering |
| Phase 1 | architect x 3-5 |
sonnet | Multi-lens idea generation (parallel) |
| Phase 2 | analyst |
sonnet | Clustering and synthesis |
| Phase 3 | critic |
opus | Quantified evaluation |
Rationalization Prevention
| Excuse | Reality |
|---|---|
| "I already know the best approach" | That's convergent thinking. Diverge first, then verify. |
| "Only one option makes sense" | Insufficient exploration. Run more lenses. |
| "Too simple for brainstorming" | Simple problems often have non-obvious better solutions. |
| "No time to brainstorm" | Bad direction costs more time than 15 minutes of ideation. |
| "Let me evaluate as I generate" | Mixing divergence and convergence kills creativity. Separate them. |
| "The first idea is good enough" | First idea = most obvious idea. Obvious ≠ best. |
Red Flags — STOP
If you catch yourself:
- Dismissing ideas during Phase 1 ("that won't work")
- Jumping to implementation before completing evaluation
- Evaluating only 1-2 ideas instead of the full set
- Skipping lenses because "we already have enough ideas"
- Selecting an idea without scoring it against the matrix
- Combining Phase 1 and Phase 3 (generating and judging simultaneously)
ALL of these mean: STOP. Return to the current phase's rules.
Integration
Pipeline position:
[brainstorm] → [deep-interview] → [plan] → [autopilot/ralph]
WHAT IF WHAT exactly HOW DO
Previous step: None (entry point for open-ended exploration)
Next step: deep-interview (for requirements clarification) or omc-plan (if requirements are already clear)
OMC delegation:
architectagents: multi-lens ideation (Phase 1)analystagent: clustering and synthesis (Phase 2)criticagent: quantified evaluation (Phase 3)exploreagent: codebase context (Phase 0)
Recommended Agent Skills
Expand your agent's capabilities with these related and highly-rated skills.
cmux-session-manager
cmux 세션 일상 관리 자동화. status(대시보드)/cleanup(정리+재조직) 수동 커맨드와 init(hook)/report(schedule) 자동화. Triggers on "cmux session", "세션 관리", "세션 정리", "cmux status", "cmux cleanup", "cmux tidy".
debug
Systematic 4-phase debugging — root cause investigation before any fix attempt. Use for any bug, test failure, or unexpected behavior. Triggers on "debug", "debugging", "why failing", "why broken", "root cause".
verify-completion
Enforce verification evidence before any completion claim. Tests, builds, and lint must be run with output shown — no claims without proof. Triggers on "verify", "verification", "done check", "completion check".
recover-sessions
Bulk recover Claude Code sessions after power loss or tmux crash. Interactive interview to determine recovery scope, layout, and execution mode. Triggers on "recover", "session recovery", "restore sessions", "power recovery".
finish-branch
Complete a development branch lifecycle — compound context, merge PR, clean up worktree and branch. Triggers on "cleanup", "worktree cleanup", "finish branch", "branch cleanup".
turbo-setup
Compound setup — issue + plan + branch + worktree + deps in one step. Triggers on "setup", "착수", "turbo-setup", "quick start".
Didn't find tool you were looking for?