Prompt engineering shapes model behavior through instructions, without retraining or infrastructure changes. Prompts are a first filter, not a guarantee; testing and monitoring are needed to catch failures they miss. section 6.2.4
Make prompts operational and testable:
- Replace vague requests such as “be fair” with specific instructions that define what the model must avoid. section 6.2.4
- Constrain response formats with allowed values, required fields, and example answers; put the instructions in the system prompt. section 6.1.3
- Evaluate prompt changes systematically, avoid relying on version-specific behavior, and pin model versions where possible. section 6.1.3
A testable fairness instruction
For a liability assessment, instruct the model: “Do not reference age, gender, or postcode.” This is more verifiable than asking the model to “be fair,” but still requires testing and monitoring. section 6.2.4 section 6.1.3


















































