The difference between a useful AI answer and a dangerous one is rarely the model. It is the brief.
What a strong brief contains
- The goal, in plain language, and the audience for the result.
- The constraints: language, framework, standards, security rules.
- The context the model cannot guess: existing schemas, naming conventions, edge cases.
- A definition of done that can be checked.
An example structure
Role: senior integration engineer
Task: map legacy order records to the new schema
Constraints: no destructive migrations; log every rejected row
Output: SQL plus a short list of assumptions to verify
Reading the output is the other half
Experts do not just write better prompts. They recognise when a plausible answer is subtly wrong. That is why we pair AI tooling with consultants who have run these systems for years.