Using a model to write or improve the prompt. Given a task description and some examples of good and bad output, models are reasonably good at producing instructions better than the ones a person wrote quickly, and there are tools that do this iteratively against an evaluation set.
Technical view
Automated prompt optimisation searches the instruction space using a model as the proposer and an evaluation set as the objective, in frameworks that treat prompts as optimisable parameters. Effective when a golden dataset exists to score against, and prone to overfitting that set when it is small. Optimised prompts are frequently unintuitive and model-specific, and must be re-optimised when the underlying model changes.
Related
Part of AI Terminology Index, section: Controlling and Extending Behaviour at Run Time.