10 prompting patterns that raise creative output

Ten prompting patterns that reliably produce more and better creative options — from forced divergence to role rotation — each with the failure it prevents.

TL;DR — Creativity from a model is a function of the prompts you give it. These ten patterns each fix a specific way a single prompt fails: it converges too early, agrees too much, or stays too generic.

Patterns that force divergence

1. Force a quota. "Give me 20 directions, not 5." Quantity first; the model's best idea is often the 14th, not the 1st.
2. Forbid the obvious. "Skip the first three ideas anyone would suggest." This cuts the generic layer.
3. Rotate the role. Ask as a skeptic, then as a child, then as a competitor. Different roles surface different ideas.
4. Change the constraint, not the goal. "How would we do this with no budget?" Constraints drive creativity.
5. Ask for the opposite. "What is the worst possible approach?" Then invert it — bad ideas are easier to generate and often point to good ones.

Patterns that improve judgment

1. Separate generation from scoring. Never ask for ideas and rankings in the same pass — the judge contaminates the generator (see the P-metrics).
2. Ask for the strongest objection. "What is the strongest argument against this?" beats "Is this good?".
3. Require a source. "For each claim, name where it comes from." Unsupported claims become visible.

Patterns that improve output

1. Give a shape. "Answer as a table with columns X, Y, Z." Structure forces the model to fill every cell honestly.
2. Iterate one dimension. Change one variable per round; you learn what actually moved the output.

These patterns are the manual version of what the DBOW loop automates. See Quick Start to run them in a pipeline.