Prompt patterns that survive

A prompt is code. It needs versioning, evaluation and tests just like everything else in your stack.

Structure beats cleverness

Clear role, task, constraints and output format outperform elaborate instructions. Put long context first or last depending on the model, and keep examples close to the task.

Evaluate or guess

Build a small set of test cases with expected behaviors. Every prompt change runs against them. Without this, prompt "improvements" are just vibes.

Frequently asked questions

How often should I revise prompts?

When models change or your evals fail — not because a new trick appeared on social media.