- Define what a correct result looks like
- Build a test set before changing the prompt
- Review the result against the source
- An illustrative quality log
- Decide whether to release
- Official reference material
- Continue with a related guide
A fluent answer can still contain an invented detail, a missing requirement or a privacy mistake. Use a small acceptance test to decide whether AI-assisted work is ready for delivery. This workflow suits drafts, summaries and structured business documents.
Define what a correct result looks like
Start with the brief and create four checks: factual accuracy, completeness, format and permitted data. For a product sheet, list the required specifications and claims that need evidence. Decide which failures block delivery; a wrong dimension matters more than a stylistic preference.
Build a test set before changing the prompt
Collect a normal case, a case with missing information and a contradictory case. Use synthetic or approved inputs. Record the expected response, including when the system should ask for clarification. Keep these examples unchanged while comparing prompt versions so an apparent improvement is not caused by easier tests.
Review the result against the source
Open every source used for a factual claim. Compare names, quantities and dates directly. Ask a second reviewer to inspect a sample without seeing your preferred conclusion. If the output invents a fact, correct the process that allowed it instead of merely fixing that sentence.
An illustrative quality log
Suppose twelve drafts are reviewed: nine pass immediately, two need minor edits and one contains an unsupported promise. Record a first-pass acceptance rate of 75%, but keep the serious failure visible separately. An average score should not hide a blocking error. These numbers illustrate a reporting method, not a benchmark for acceptable quality.
Decide whether to release
Deliver only when blocking issues are resolved and the acceptance checks are complete. Keep the prompt version, source list and reviewer note with the final file. Re-run the same tests when changing the model or workflow; improvement on familiar cases is not proof that all future inputs are safe.
Official reference material
Continue with a related guide
Apply one useful step, measure the result, and continue with the most relevant guide.
