Skip to main content

Improve and Review Prompts

Improve a prompt by comparing its output with a known expectation, changing one missing requirement, and testing again.

The Concept​

Do not respond to a weak result by adding random instructions. Diagnose the failure: missing source, vague objective, wrong format, unsupported assumption, or unsafe action. Update the controlled instruction in Agent Prompts, test representative cases, and inspect any external action in Tool Executions.

Example​

Weak output: a risk review lists claims without evidence.

Prompt change:

For every risk, add Evidence, Source page, Confidence, and Missing information.
If no supporting passage exists, mark the risk Unverified.

Test the revised prompt with a normal case, a missing-evidence case, and a conflicting-source case.

Checklist​

  • The expected result can be described before the run.
  • The failure is assigned to one clear cause.
  • One controlled instruction changes at a time.
  • Known good, missing, and conflicting cases are tested.
  • Only reviewed prompt versions are reused or shared through the Library.

Next, apply the same evidence and review model at larger scale in Enterprise AI Patterns.

Last updated on