diverse-ideas

name

diverse-ideas

description

Generate highly diverse ideas using research-backed prompting techniques from the Wharton paper "Prompting Diverse Ideas" (Meincke, Mollick, Terwiesch, 2024). Use this skill whenever the user wants to brainstorm, ideate, generate options, explore creative directions, come up with names, strategies, solutions, features, products, content topics, or any task where variety and novelty matter. Also trigger when users say things like "give me ideas", "brainstorm", "what are some options", "help me think of", "creative solutions for", "explore possibilities", or any request where generating a diverse set of alternatives would be valuable — even if the user doesn't explicitly ask for "diverse" ideas.

Diverse Ideas Generator

Based on the Wharton paper "Prompting Diverse Ideas" (Meincke, Mollick, Terwiesch, 2024) and 25 benchmarked experiments. See LEARNINGS_DIVERSITY.md for full data.

Core insight

More agents with different starting perspectives beats a single agent with a better prompt. Our benchmarks showed that 21 isolated agents (1 CoT + 10 dynamically generated personas + 10 creativity techniques) produce ideas with a 30.7% surprise rate, vs 15.3% for a single CoT agent and 21.4% for vanilla prompting. Each agent runs in a fresh context with no knowledge of the others, preventing the anchoring that kills diversity.

The foundation is the paper's Chain-of-Thought (CoT) prompt: generate 100 short titles (breadth), self-critique for boldness (diversity), then elaborate (depth). Every agent uses a variant of this three-step pattern. But the real diversity comes from the variety of starting perspectives, not from any individual prompt being clever.

The process

Step 1: Refine the problem (if ambiguous)

If the prompt lacks 2+ of: target audience, domain/industry, constraints, success criteria, or starting context, ask 1-5 focused questions. If it's already specific or the user says "just go", proceed.

Step 2: Choose the mode

Present the user with a quick choice:

• Quick (default for simple requests): 2 agents. Pure CoT + Informed CoT. ~28 ideas, ~4 min. Good enough for most brainstorming.
• Deep (for serious ideation, high-stakes decisions, or when the user asks for more diversity): 21 agents. CoT + 10 personas + 10 techniques. ~28 curated ideas from ~300 raw candidates, ~5 min. Roughly 2x the surprise rate of Quick mode.

If the user doesn't specify, use Quick for casual requests ("give me some ideas for...") and Deep for explicit ideation requests ("brainstorm", "I need diverse ideas", "explore all options").

Step 3: Generate personas (Deep mode only)

Spawn a separate agent to generate 10 personas for the specific problem:

Generate 10 diverse personas for brainstorming ideas about: [user's problem]

Each persona should be a specific professional role FAR from the problem domain,
whose daily work involves structurally similar challenges seen from a completely
different angle.

Selection criteria:
- Domain distance: expertise far from the problem domain
- Specificity: concrete identity with a worldview, not generic labels
- Cognitive diversity: spread across thinking modes (spatial, systems, embodied,
  aesthetic, quantitative, social, temporal, adversarial)
- Independence: no two personas should be adjacent

Return ONLY valid JSON:
[{"persona": "Role Name", "description": "One sentence about their perspective."}, ...]

Step 4: Launch all agents in parallel

Spawn all agents simultaneously using the Agent tool. Each runs in a fresh, isolated context.

CoT Explorer (always included, both modes):

You are generating ideas for: [user's problem]
Context: [any relevant context]
Constraints: [any constraints]

Follow these steps. Do each step, even if you think you do not need to.

Step 1: Generate a list of 100 ideas (short title only).
Cast an extremely wide net across different domains, user needs, form factors,
and unconventional angles. Include weird, surprising directions.

Step 2: Go through the list and determine whether the ideas are different and bold.
Modify as needed to make them bolder and more different. No two ideas should
be the same. This is important!

Step 3: Select the 30 most novel and diverse ideas. For each, give it a name
and a brief description (1-2 sentences). Prioritize range over polish.

Return ONLY your final curated list of 30 as a numbered list.

Informed CoT Explorer (included in Quick mode, optional in Deep):

Same as CoT Explorer but with a research step prepended:

STEP 0 — RESEARCH: Before generating ideas, use WebSearch to research this topic.
Run 2-3 searches about what already exists, what pain points users report,
and what trends are emerging. Take brief notes, then use your research as a
springboard (not a constraint) for ideation.

[Then Steps 1-3 same as CoT Explorer]

Persona Agents (Deep mode, one per persona):

You are a [persona]. [description]

You are generating ideas for: [user's problem]

Step 1: Generate 50 ideas (short title only) from your unique perspective.
What patterns from your world apply here? What would you try that nobody
in [the problem's domain] would think of?
Step 2: Review and push bolder. Are these truly different from what an
ordinary person would suggest? Revise any that feel safe.
Step 3: Select your 15 most novel ideas with brief descriptions.

Return ONLY your final 15 as a numbered list.

Creativity Technique Agents (Deep mode, one per technique):

You are generating ideas for: [user's problem]
Use the [technique] creativity technique: [instructions]

Step 1: Apply the technique to generate 50 raw ideas (short title only).
Step 2: Push bolder. Remove variations of the same concept.
Step 3: Select your 15 most novel ideas with brief descriptions.

Return ONLY your final 15 as a numbered list.

The 10 techniques and their instructions:

1. SCAMPER: Apply each operation (Substitute, Combine, Adapt, Modify, Put to other use, Eliminate, Reverse). 7+ ideas per operation.
2. Constraint Removal: List 5 biggest constraints. Remove each, generate 10 ideas. Adapt impossible ideas back toward reality.
3. Analogical Transfer: 5 different domains solving structurally similar problems. 10 ideas per domain by transferring patterns.
4. Random Entry: 5 random unrelated stimuli. Force unexpected connections. 10 ideas per stimulus.
5. First Principles: Decompose to fundamentals. Forget convention. Rebuild 50 solutions from scratch.
6. Reverse Brainstorming: Brainstorm 20 ways to make the problem worse. Flip each into a solution.
7. Worst Possible Idea: Generate 25 deliberately terrible ideas. Find the useful kernel in each and transform it.
8. Six Thinking Hats: Ideas from 6 perspectives (factual, emotional, cautious, optimistic, creative, process). 8+ per hat.
9. Biomimicry: 5 biological systems solving analogous problems (ant colonies, immune systems, mycelium, coral reefs, migration). 10 ideas per system.
10. Time Travel: Ideas as if the problem existed in 5 eras (ancient Rome, medieval, 1950s, 2050, 500 years from now). 10 per era.

Step 5: Merge and curate

Once all agents return, merge their combined ideas:

1. Deduplicate: Remove ideas that are the same concept, keeping the more interesting framing.
2. Cluster by theme: Group remaining ideas into natural clusters.
3. Select ~25 for maximum diversity across clusters, prioritizing:
   • Ideas unique to one agent (these are the most valuable)
   • A mix of pure-novelty and research-grounded ideas
   • Surprising or counterintuitive ideas
   • A balance of practical and ambitious
4. Present to the user:

## Ideas for [topic]

### [Theme 1]
1. **Idea name** — Brief description
2. **Idea name** — Brief description

### [Theme 2]
...

### Wild Cards
N. **Idea name** — Brief description (deliberately unconventional)

Include a "Wild Cards" section with 2-3 of the most provocative ideas.

Step 6: Go deeper (if the user wants)

After presenting, offer: "Want me to develop any of these further, or run another round focusing on a specific direction?"

Anti-patterns (what hurts diversity)

• Never seed agents with examples or previous ideas. The paper proved this reduces diversity (cosine similarity jumped from 0.377 to 0.428).
• Never tell agents about similarity between ideas. The model games the metric instead of exploring.
• Never generate all ideas in one context. Shared history causes anchoring. Each agent must be isolated.
• Never use domain experts as personas. The user already has that lens. Pick personas whose daily work involves structurally similar problems in completely unrelated fields.