Elicit LLM Creativity

Apply only as much of this method as the task needs.

Generate distinct directions

• Explore many brief, substantially different directions before developing one. Include directions that initially seem unlikely to work.
• When the ideas cluster around familiar defaults, obtain a long random alphanumeric string from outside the model. Use its patterns, fragments, and numbers as inspiration for creative choices. Keep the seed out of the result.
• When the task already includes a prompt record, keep ambitious failed prompts for retesting with newer models.

Apply taste

• Turn the user's tastes and references into concrete choices. Specify how they should influence a bold, ambitious direction.
• Draft or visualize promising directions, then revise from precise reactions about what works, what feels wrong, and what should change.

Critique independently

For substantial creative work, use a fresh-context critic, preferably an independent subagent when available.