Low Between-Strategy Overlap

Different prompting strategies produce genuinely different ideas, not just the same ideas expressed differently.

When comparing the first 50 ideas from different strategies (Steve Jobs persona, emotional appeals, HBR techniques, etc.), the overlap is surprisingly small. Each strategy appears to access a different region of the idea space.

This has a practical implication: hybrid brainstorming, running multiple strategies and combining their outputs, is more effective than running one strategy longer. Rather than generating 100 ideas with Chain-of-Thought, you might get more diversity from 25 ideas each from four different strategies.

The finding also suggests that prompt engineering doesn’t just improve output quality within a fixed distribution. It shifts which distribution the model samples from. The same underlying model contains multiple “idea generators” that can be activated by different prompts.

Related: 05-atom—idea-space-as-landscape, 05-atom—cot-diversity-effect, 05-molecule—prompt-engineering-for-diversity