A useful prompt tells the system what to do and gives a reviewer a way to judge whether it did it. Elaborate role descriptions do not replace clear inputs, constraints and evidence requirements.
Start with the decision or deliverable
Replace “analyze our business” with a bounded request such as “summarize changes in monthly revenue by product using the supplied table, and identify which changes need investigation”. State the audience and the output length only when those constraints serve the task.
Define whether the system should summarize supplied material, calculate values or propose possible explanations. Those are different operations with different verification needs. If new research is required, specify which sources are acceptable and how claims must be supported.
Set an evidence boundary
Tell the system what information it may use and how to handle gaps. For a closed-source task, ask it to rely on the supplied material and mark unsupported questions as unresolved. For research, require actual source links and dates, then open them yourself.
A rule against inventing citations is a useful instruction, not an enforcement mechanism. The reviewer must still confirm that a cited document exists and supports the relevant claim. Keep that distinction visible in the workflow.
Use a structured output
For a management summary, a practical structure is: observed change, supporting calculation, possible explanation, missing evidence and proposed next check. This prevents a speculative explanation from appearing in the same column as a verified number.
| Weak request | More inspectable request |
|---|---|
| Tell me why sales fell | Calculate the change from the supplied data; list possible explanations separately and identify evidence needed to test them |
| Write the best customer reply | Draft a reply using only the approved policy; flag exceptions for human review |
| Give an accurate report | Show sources and units for each material figure and list unresolved questions |
Evaluate changes one at a time
Keep a version of the prompt and test it against the same representative cases. If you change the input, prompt and tool simultaneously, you cannot easily tell what caused the outcome. Record review time as well as output quality.
The Prompt Lab on this site assembles a brief with task, audience, context, format and verification instructions. It is a template builder, not a model. Before using the generated brief elsewhere, replace placeholders with approved material and check whether the chosen tool is appropriate for that data.