Describe the input, intended result and review criteria when writing an AI request. These details give the task a clear scope and provide a reference for checking the response.
Anthropic’s prompting guidance recommends clear instructions and examples. The method below puts those ideas into a short task description with concrete review criteria.
A request is easier to review when it describes what a usable result looks like. For a document edit, that might mean preserving every date, meeting an agreed length and listing unresolved questions. These details give the writer and reviewer a shared reference.
Specify what the result is for
Instead of improve this document, try: edit this supplied notice for residents, preserve every date and instruction, keep it within the stated length and list any ambiguity separately. That gives the writer and reviewer something specific to inspect.
Identify which source material is authoritative. Ask the system to list any missing information as questions for review.
Use four fields
| Field | Question |
|---|---|
| Input | Which material should the system use? |
| Output | What must be returned, and for whom? |
| Constraints | What must remain, change or be excluded? |
| Acceptance | Which checks will make the result usable? |
Keep examples relevant and label hypothetical material. An example should clarify the requested result without introducing an unsupported fact the system might repeat.
Check the requirements and readability
Check whether the answer used the supplied input, preserved important constraints and completed the requested work. A polished paragraph can still omit a date or change a requirement.
If the request keeps producing unusable results, revise one part of the task and compare again. Keep a record of actual failures. Keep examples of the results and revisions so the request can be assessed across the work it is intended to support.
Use a complete fictional task specification
A request for editing a residents’ notice could be written as follows:
Edit the supplied notice for residents. Preserve all dates, locations, actions and contact details. Keep the revised notice within 180 words. Return the revised text and a separate list of unclear points. Treat missing information as questions for review.
The 180-word limit is an example constraint, not a general target. It gives this task a measurable requirement alongside meaning and readability.
Translate the request into review checks
Prepare a short checklist before inspecting the response: all source dates retained, required actions preserved, length within the stated limit, unclear points listed separately and no unsupported additions. Compare the result with that checklist and the original notice.
If the task needs revision, change the part that the result exposed. A missing audience description may need a clearer audience; a repeated omission may need an explicit preservation requirement or a smaller unit of work. Keep the revised request with the example so the next attempt can be compared.
For a larger project, identify the deliverable and completion conditions for each stage. A prepared result, a reviewed result and a published result can then be tracked clearly, with the relevant decision at each transition.
Keep the source and revised text side by side for a focused check of names, numbers and conditions. Then read the revision for flow and clarity. Splitting the review into those two passes gives both meaning and readability their own attention.
Sources consulted
- Prompting best practices — Anthropic; consulted 3 November 2026.