Define a task, not a personality
“Be an expert” is less useful than a concrete instruction about what to do. State the objective, audience, input and expected output. For example: summarize the supplied maintenance note for a new technician, in five bullets, preserving every measurement and marking ambiguous instructions. This creates criteria you can actually inspect.
Give boundaries and examples
Specify whether outside knowledge is allowed, how unknowns should be handled, and what must remain verbatim. A short example can clarify format. Keep examples representative: a model may imitate their mistakes or assumptions. Quote source material separately from instructions, especially when working with untrusted documents that could contain commands aimed at the assistant.
Evaluate representative failures
Build a small test set that includes ordinary inputs, missing information, conflicting dates and adversarial instructions inside documents. Keep expected facts or a scoring rubric. Compare versions on the same set. Do not select only your best-looking answer; a prompt that succeeds once may fail on a different document or run.
Keep a human review step
Check unsupported claims, omitted qualifications, source coverage and adherence to format. For a drafted email, review the recipient, commitments and attachments before sending. For a calculation, independently recompute the result. A numeric score is useful only if the criteria behind it are meaningful and consistently applied.
Try it yourself
Write a prompt for extracting dates from a fictional meeting note. Require a quoted supporting passage and “not specified” where a date is missing. Test a note with two conflicting dates.
What to look for
A strong result preserves both dates with their source text and flags the conflict. It should not invent a resolution or silently choose one date.
Sources & further reading
Independent educational material. Examples are simplified and do not replace checking the documentation for the model or service you use.