Designing and Testing Robust Prompts

A professional prompt is a transfer of work. Start with the end result and then add the relevant context, desired output format, and necessary boundaries. Indicate which source information is leading and what should be done in case of missing data.

A role can guide perspective or jargon, but it does not prove expertise. Therefore, it is better to use concrete assessment criteria rather than an impressive job title. Conclude important tasks with a control instruction: report missing information, unsupported claims, and deviations from the assignment separately.

• Result
• Relevant context
• Usable format
• Real boundaries
• Post-check

With example-driven instruction, you show input and desired output. This is especially helpful for classification, fixed formats, tone of voice, and exceptions. Choose examples that together cover the normal cases and important boundaries.

Explain what the model should adopt from an example: structure, level of detail, criterion, or tone. Do not automatically reuse names, facts, or incidental wording. Include at least one unusual or incomplete case and define the correct behavior, for example 'insufficient information' instead of guessing.

• Representative examples
• Name desired feature
• Include edge case
• No unwanted fact copying

For professional work, you want to be able to trace what an outcome is based on. Therefore, ask for source extractions, assumptions, calculations, decision criteria, and a concise justification. You can check such artifacts without pretending that a generated explanation fully reveals the actual internal reasoning process.

Choose intermediate steps that locate errors. In a comparison, these are, for example, criteria, source values, missing data, and scores. In a text check, these are claims, evidence, and suggested corrections. A long line of thought is not evidence; verifiable data and repeatable tests are.

• Source extraction
• Assumptions separately
• Calculation or criterion
• Concise justification
• Remaining uncertainty

A prompt that handles one example correctly is not yet a reliable workflow. Create a test set with representative input, a difficult case, missing information, conflicting sources, and possibly an inappropriate instruction in the source material. Define in advance for each test what minimally correct behavior is.

Do not use real sensitive data when fictional examples suffice. Evaluate both content and safety: does the output follow the schema, are facts not being made up, does the source content remain subordinate to the task, and does the workflow stop where it should? Save failed examples as regression tests.

• Normal case
• Incomplete case
• Contradiction
• Out of scope
• Unreliable source instruction

Do not compare prompt versions based on feeling or on a single random answer. Use the same test set and the same evaluation criteria. Preferably change only one factor, such as source instruction, output schema, or error handling. That way you will know which adjustment likely made the difference.

Record version number, change, reason, tests, and known limitation. Measure not only quality but also correction work, consistency, and required time. A version with a higher average score can still be unsuitable if it handles one critical edge case incorrectly.

• Same test set
• One main change
• Criteria before test
• Critical errors separately
• Version log