Designing the instructions given to a language model — role, task, constraints, output format, examples — so that it produces a usable result consistently rather than occasionally.
The same model can be reliable or useless depending on how it is asked. In a product the prompt is code: it is versioned, reviewed, and tested, because a careless edit changes behaviour for every user.
We keep prompts in the repository under review, pin model versions, and maintain a test set of real inputs with expected outputs, so a prompt change is validated before it ships rather than discovered in production.