Prompt Engineering
Make the request inspectable.
Good prompting is not a hunt for magic words.
Prompt engineering is the design of instructions for a specific task. A useful prompt makes the work easier to produce, review, and improve. When important details are missing, a model may fill the gaps incorrectly.
Five parts of a useful prompt
- Role, when useful: Name a perspective or audience, but do not treat a role as proof of expertise.
- Objective: State the outcome and what success means.
- Task context: Provide relevant facts, sources, and constraints.
- Deliverable: Describe the work and required format.
- Audit request: Ask for sources, assumptions, missing information, and checks against your criteria.
An audit request is evidence for review.
It does not prove the answer is correct. For consequential work, compare important claims with the actual sources.
Make the source boundary visible.
Help a first-year student explain the assigned concept. Use only the chapter I provide. Explain it in plain language, give one practical example, and ask three questions that test understanding. Cite the relevant section for each important claim. Mark anything the chapter does not support as UNVERIFIED.
You define the purpose and evidence standard.
You also decide when the result is good enough, and when it requires outside verification.
Rewrite one real request.
Use the five parts. Test it on three representative examples, keep one failure case, and revise the prompt.
The real skill is not making AI sound certain. It is making the request and result easier to inspect.
Next: Context EngineeringSource check: Google AI, Prompt design strategies.