Introduction
Goal
Generate a high volume of genuinely diverse solution ideas, then organise them into actionable themes.
Method
Frame the problem as How Might We questions, spawn parallel agents with different diversity primes, then deduplicate and group the ideas inline. Each prime is either context-blind (for divergent thinking) or context-aware (for grounded thinking). The whole workflow runs in a single pass. The user sees nothing until the final themed output.
Success Criteria
- Volume — Many ideas per problem area
- Diversity — Different approaches, not variations of one
- Problem-anchored — Each idea addresses a specific HMW question
- Actionable — Concrete enough to evaluate
- Novelty — Not just restated existing solutions
Workflow
1. Frame
- From the user's problem or goal, generate How Might We questions:
- One HMW per problem area (or group related areas)
- Format: "How might we [action] so that [outcome]?"
2. Select primes
- Pick 2 primes. One context-blind, one context-aware. Use more only when the user asks for extra breadth. Genuine diversity comes from different context or lenses:
| Type |
Examples |
| Context-based |
No context (naive), full context (strategic) |
| Lens-based |
Minimalist, enterprise, developer-focused, end-user-focused |
| Analogy-based |
Far-field (how unrelated domains solve the same underlying structure) |
- Context-blind primes derive value from not knowing (naive, outsider, cross-domain, minimalist, far-field). Ignorance of current reality is the feature.
- Context-aware primes derive value from grounding in reality (developer-focused, pragmatist, strategic, enterprise). Without project context they produce generic ideas.
- For each context-aware prime, identify what context to gather. Be specific (e.g., "read src/auth/ to understand current auth flow") rather than gathering everything.
3. Gather context
- Gather the identified context for context-aware primes. Read the few files that matter directly. Use an Explore agent only when the relevant context is too large to read directly.
- Skip this step if all primes are context-blind.
4. Generate
- For each prime, spawn a blind-reasoner agent (Haiku). Name each after its prime.
Context-blind prompt:
You are brainstorming solutions from a {prime} perspective. Work ONLY with what is given below.
HMW QUESTIONS:
{hmw_questions}
Generate 5 solution ideas per HMW question. Be concrete and specific. Think from your assigned perspective — a {prime} thinker would approach this differently from other perspectives.
Context-aware prompt:
You are brainstorming solutions from a {prime} perspective. Work ONLY with what is given below.
PROJECT CONTEXT:
{gathered_context}
HMW QUESTIONS:
{hmw_questions}
Generate 5 solution ideas per HMW question. Be concrete and specific. Think from your assigned perspective — a {prime} thinker would approach this differently from other perspectives. Use the project context to ground your ideas in what actually exists, but don't limit yourself to incremental improvements.
- Run all agents in parallel.
5. Synthesise
- This step runs inline. Spawn no further agents.
- Merge duplicate and near-duplicate ideas across agents.
- Where two ideas from different agents combine into something stronger, add the hybrid.
- Group the ideas into at most 7 themes by the dimension of the problem each addresses. Avoid grouping by surface similarity.
Final output
- Present a single message. The HMW questions, one line on the primes used, and the ideas grouped by theme.
- If the user wants more, offer a build-on round (resume each generator with the other agents' strongest ideas) or a full group-ideas pass. Run these only on request.
Related skills
- clarify-framing — If the problem isn't clear, run clarify-framing first.
- breakdown-problem — If the problem is too large, break it down first and brainstorm per sub-part.
- group-ideas — For a more rigorous grouping than the inline pass.
1---2name: brainstorm3description: Generates diverse solution ideas using parallel agents with different perspectives.4---56# Introduction78## Goal9Generate a high volume of genuinely diverse solution ideas, then organise them into actionable themes.1011## Method1213Frame the problem as How Might We questions, spawn parallel agents with different diversity primes, then deduplicate and group the ideas inline. Each prime is either context-blind (for divergent thinking) or context-aware (for grounded thinking). The whole workflow runs in a single pass. The user sees nothing until the final themed output.1415### Success Criteria1617- **Volume** — Many ideas per problem area18- **Diversity** — Different approaches, not variations of one19- **Problem-anchored** — Each idea addresses a specific HMW question20- **Actionable** — Concrete enough to evaluate21- **Novelty** — Not just restated existing solutions2223## Workflow2425### 1. Frame26- From the user's problem or goal, generate How Might We questions:27 - One HMW per problem area (or group related areas)28 - Format: "How might we [action] so that [outcome]?"2930### 2. Select primes31- Pick 2 primes. One context-blind, one context-aware. Use more only when the user asks for extra breadth. Genuine diversity comes from different context or lenses:3233| Type | Examples |34|------|----------|35| Context-based | No context (naive), full context (strategic) |36| Lens-based | Minimalist, enterprise, developer-focused, end-user-focused |37| Analogy-based | Far-field (how unrelated domains solve the same underlying structure) |3839- **Context-blind** primes derive value from not knowing (naive, outsider, cross-domain, minimalist, far-field). Ignorance of current reality is the feature.40- **Context-aware** primes derive value from grounding in reality (developer-focused, pragmatist, strategic, enterprise). Without project context they produce generic ideas.41- For each context-aware prime, identify what context to gather. Be specific (e.g., "read src/auth/ to understand current auth flow") rather than gathering everything.4243### 3. Gather context44- Gather the identified context for context-aware primes. Read the few files that matter directly. Use an Explore agent only when the relevant context is too large to read directly.45- Skip this step if all primes are context-blind.4647### 4. Generate48- For each prime, spawn a blind-reasoner agent (Haiku). Name each after its prime.4950**Context-blind prompt:**51```52You are brainstorming solutions from a {prime} perspective. Work ONLY with what is given below.5354HMW QUESTIONS:55{hmw_questions}5657Generate 5 solution ideas per HMW question. Be concrete and specific. Think from your assigned perspective — a {prime} thinker would approach this differently from other perspectives.58```5960**Context-aware prompt:**61```62You are brainstorming solutions from a {prime} perspective. Work ONLY with what is given below.6364PROJECT CONTEXT:65{gathered_context}6667HMW QUESTIONS:68{hmw_questions}6970Generate 5 solution ideas per HMW question. Be concrete and specific. Think from your assigned perspective — a {prime} thinker would approach this differently from other perspectives. Use the project context to ground your ideas in what actually exists, but don't limit yourself to incremental improvements.71```7273- Run all agents in parallel.7475### 5. Synthesise76- This step runs inline. Spawn no further agents.77- Merge duplicate and near-duplicate ideas across agents.78- Where two ideas from different agents combine into something stronger, add the hybrid.79- Group the ideas into at most 7 themes by the dimension of the problem each addresses. Avoid grouping by surface similarity.8081## Final output82- Present a single message. The HMW questions, one line on the primes used, and the ideas grouped by theme.83- If the user wants more, offer a build-on round (resume each generator with the other agents' strongest ideas) or a full group-ideas pass. Run these only on request.8485## Related skills86- **clarify-framing** — If the problem isn't clear, run clarify-framing first.87- **breakdown-problem** — If the problem is too large, break it down first and brainstorm per sub-part.88- **group-ideas** — For a more rigorous grouping than the inline pass.