Prompt Engineering for Diversity
The Problem
When you ask an LLM to brainstorm, it produces ideas that cluster in a narrower region of possibility space than human groups would explore. The ideas are competent and well-formed, but they lack variance. In innovation contexts where the value of your best idea matters more than your average idea, this clustering limits results.
Why It Happens
LLMs learn statistical patterns from human-generated text. They’re optimized to produce plausible, expected outputs. For routine tasks, this is a feature. For brainstorming, it’s a bug, the model gravitates toward the well-trodden center of idea space rather than the unexplored edges where novel solutions live.
What Works
Chain-of-Thought Prompting (Most Effective) Break ideation into sequential steps:
- Generate short titles only (forces breadth before depth)
- Review and modify for “boldness and difference”
- Expand into full descriptions
This approach nearly matches human group diversity (0.255 vs 0.243 cosine similarity) compared to baseline prompting (0.377).
Hybrid Strategy (Maximizes Coverage) Different prompting strategies access different regions of idea space. Rather than generating many ideas with one approach:
- Run 4-5 different strategies for shorter bursts
- Combine outputs and de-duplicate
- Select diverse representatives from the combined pool
Persona Prompts (Simple Improvement) Simple personas (“You are Steve Jobs”) outperform elaborate methodology instructions. The model has internalized “creative person” as a pattern; persona labels surface it.
What Doesn’t Work as Expected
- Emotional appeals (“I’ll get fired if these aren’t creative!“)
- Tipping promises (“I’ll tip $200 for novel ideas”)
- Seeding with example ideas (slightly increases clustering)
- Including cosine similarity feedback (no improvement)
Practical Constraints
- Benefits concentrate in the first ~500 ideas
- All strategies converge after ~750 ideas as the accessible space exhausts
- Total unique idea capacity: ~3,700 (baseline) to ~4,700 (CoT)
When to Apply
Use these techniques when diversity is the goal, early-stage brainstorming, exploring unfamiliar problem spaces, generating options before converging. For tasks where consistency matters (documentation, analysis, summarization), the baseline approach is fine.
Related: 05-atom—ai-diversity-deficit, 05-atom—uniform-confidence-problem