BeCreative

Use Verbalized Sampling, which asks the model to internally generate 5 diverse options with probabilities below 0.10 each, then output the single best one. Research shows this increases creative diversity by 1.6-2.1x compared to standard prompting.

Verbalized Sampling is a technique from Zhang et al. (2024) that forces exploration of low-probability creative options, circumventing typicality bias in alignment training. Combined with extended thinking, it improves both diversity and output quality by about 25.7%.

Use the TechnicalCreativityGemini3 workflow, which shells out to the llm CLI with a structured prompt covering problem, constraints, and success criteria. Run the preflight check 'which llm && llm models list | grep -i gemini-3' first to confirm the CLI and model are available.

The core creativity workflows run entirely through prompt templates and extended thinking with no dependencies. Only the TechnicalCreativityGemini3 workflow requires the llm CLI with the llm-gemini plugin and a configured Gemini API key.

Use TreeOfThoughts for complex problems with multiple constraints where you need to explore 3-5 fundamentally different approaches and their sub-branches. Standard brainstorming suits simpler idea generation where a single round of diverse options is enough.

These techniques improve diversity but still depend on the model's training distribution, so truly novel domains may see limited gains. They also consume more tokens due to internal option generation, and quality still requires human evaluation for high-stakes creative work.