fatby/diverse-ideation

--- name: diverse-ideation description: Generate maximally diverse sets of ideas on any topic using Chain-of-Thought prompting to avoid LLM mode collapse. Use when brainstorming business ideas, product concepts, strategies, solutions, or any creative task where you need high-variance output with many genuinely different ideas rather than variations on the same theme. Always followed by a separate assessment/pruning step (not part of this skill). --- # Diverse Ideation Generate a maximally diverse set of ideas on any topic by spawning a subagent with Chain-of-Thought (CoT) instructions. Based on research showing CoT prompting achieves near-human-level idea diversity (cosine similarity 0.255 vs humans at 0.243, compared to naive prompting at 0.377+). ## Core Mechanism CoT works by breaking ideation into distinct micro-steps: 1. Generate short titles only (fast, divergent) 2. Review for similarity — actively modify to make bolder/more different 3. Expand into full descriptions This explicit review-and-revise step is what prevents mode collapse. The model literally checks its own output for sameness before elaborating. ## How to Use ### 1. Distill Context Into a Brief Before spawning, compress all relevant context into a concise brief. Include: - Domain/topic: What are we ideating about? - Constraints: What's in scope and out of scope? (budget, technical feasibility, resources, timeline) - Target audience/market: Who is this for? - Known solutions to avoid: What already exists that we don't want to reinvent? - Evaluation criteria preview: What will "good" look like in the pruning step? (helps the subagent calibrate without over-filtering) - Entropy sources: Any specific real-world data, trends, signals, or examples to seed from (critical for avoiding generic output) The more specific and grounded the context, the more diverse AND relevant the output. Generic briefs produce generic ideas. ### 2. Spawn a Subagent Use sessions_spawn with a prompt following this template: You are generating a maximally diverse set of ideas about [TOPIC]. Context: [DISTILLED BRIEF — domain, constraints, audience, known solutions, entropy sources] Follow these steps. Do each step, even if you think you do not need to. Step 1: Generate a list of [N] ideas as short titles only (one line each). Cover as many different angles, mechanisms, and approaches as possible. No two ideas should address the same need in the same way. Step 2: Review your list. For each idea, check: is it genuinely different from every other idea in primary mechanism AND target need? If two ideas feel similar, replace one with something from a completely different direction. Be aggressive about this — similarity is the enemy. This step is critical. Step 3: For each idea, write a short description (2-4 sentences) covering: what it is, how it works, who it's for, and why it's different from obvious alternatives. Format your final output as a numbered list with "Title: Description" format. Remember: The goal is HIGH RECALL of the idea space, not filtering for quality. Bad-but-different ideas are better than good-but-similar ideas. Quality filtering happens later. ### 3. Key Parameters - N (number of ideas): 30-50 is the sweet spot per session. Beyond ~100, exhaustion effects kick in and ideas start repeating regardless of prompting. - Model: Use the best available model. CoT benefits more from capability than speed. - Multiple sessions: For maximum coverage, run 2-3 sessions with slightly different framings or entropy sources, then combine. Cross-session overlap is typically low (paper finding: between-pool similarity is much lower than within-pool). ### 4. Boost Diversity Further When you need even more variance: - Vary the framing across sessions: different personas, different constraint emphasis, different starting categories - Inject entropy sources: feed in specific data points, trends, or examples from real-world research — this grounds the ideas and prevents generic output - Use analogical thinking prompts: "What would [unrelated industry] do to solve this?" "How would [famous innovator] approach this?" - Spawn parallel sessions with different seeds and merge results (works better than sequential ideation) ## Example Use Cases - Product ideation: Generate 50 distinct product concepts for the "smart home" space - Content planning: Brainstorm 40 different angles for a blog series on AI safety - Business strategy: Ideate 30 different growth tactics for a B2B SaaS company - Creative writing: Generate 20 wildly different plot concepts for a sci-fi novel - Problem solving: Come up with 25 different approaches to reduce carbon emissions in cities ## Related Skills - Idea assessment/pruning: Use after diverse ideation to filter, rank, and select the best ideas (not included in this skill) - SWOT analysis: Evaluate ideas from multiple perspectives - Risk analysis: Identify potential pitfalls - Feasibility check: Assess technical/business viability ## References - Research paper: "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (Wei et al., 2022) - Follow-up study: "CoT Improves Idea Diversity in LLM Brainstorming" (Chen et al., 2024) - OpenClaw documentation: sessions_spawn tool - Anthropic skills specification: SKILL.md format