Prompt Engineering

A useful prompt makes the task, relevant information, constraints, and expected output clear, then is evaluated against examples of success and failure.

주요 시사점

• Define the task and success criteria before optimizing the wording.
• Use representative test cases, including missing or conflicting information.
• Prompt instructions support reliability but do not replace validation or security controls.

Start with the outcome rather than a special phrase. Decide what the model must produce, which information it may use, and how you will check the result. If you cannot distinguish a good answer from a bad one, changing the prompt can give the appearance of progress without improving the task. A practical prompt separates instructions from input data, supplies the context needed for the task, and specifies the output format. Examples can clarify an ambiguous format or distinction. Do not assume that a persona such as 'expert researcher' gives the system real expertise or access to evidence that was never provided. Build a small evaluation set containing ordinary inputs and difficult cases: missing information, conflicting statements, unusual formatting, and requests outside the intended scope. Change one important part of the prompt at a time and compare the outputs. Record both improvements and regressions. Prompting has limits. It cannot make unavailable information appear, guarantee factual accuracy, or replace access controls. For sensitive workflows, validate outputs, restrict tool permissions, and decide which actions need human review. Treat instructions contained inside untrusted documents as data to examine, not authority to change the task.

1. Vague request: 'Summarize this event.' This does not say which information matters or how to handle omissions.
2. Testable request: 'Extract the event name, start time, and end time from the note below. Return only those three fields. Use null for anything not stated. Do not infer an end time.'
3. Test with the invented note 'Model Workshop starts at 10:00.' Check that the result includes Model Workshop, 10:00, and a null end time. Then add a conflicting time and decide in advance how that case should be handled.

You now have an explicit task and a checkable expected result. Run the test against the model you plan to use; a well-written prompt is not itself proof that the model passes.