AI-assisted practical guide. Examples are hypothetical; proposed workflows are editorial suggestions.
“Make this better” leaves an AI tool to guess what success means. A stronger prompt defines the input, expected output, exclusions, and checks that a reviewer can actually perform.
Describe an observable result
State the intended audience and task. Specify required sections or fields only where they matter. List facts that must remain unchanged and information the model must not invent. Explain how missing inputs should be handled.
Keep the acceptance criteria independent of the model's self-assessment. Asking it to declare success is not the same as checking the result.
Try a hypothetical formatting task
Suppose a fictional task asks for notes to become a one-page briefing. “Done” might mean every supplied decision appears, unresolved questions remain labeled, no new factual claims are added, and the requested headings are present. Those checks are more useful than demanding that the result be “professional.”
Review the output against each criterion. If the criteria conflict, resolve the conflict rather than hoping the model chooses your preference.
Refine from actual failures
Save examples where the output missed the target. Adjust the prompt to address the failure without adding unrelated rules. Rerun a small fixed task packet after meaningful changes. The goal is a clearer working agreement between request and review, not an increasingly long prompt that nobody can understand or maintain.