chatgpt
Write prompts you can check
A useful prompt defines the task, supplies the evidence, and makes the answer easy to check. Practice with a small example before adding a long list of roles or instructions.
Start with a result you can verify
“Summarize this” leaves the output unclear. For a meeting handoff, specify the fields and what should happen when the notes are incomplete.
This example uses invented notes:
Ari will review the draft on Thursday.
The budget needs approval; no owner was named.Give the model a clear contract
Create a handoff from the notes below.
Return two sections: Assigned actions and Open questions.
Assigned actions must include task, named owner, due date,
and a supporting quote.
Only include details present in the notes.
Write "not specified" for a missing due date.
Put tasks without a named owner under Open questions.
Do not invent approvals, deadlines, or decisions.
Notes:
Ari will review the draft on Thursday.
The budget needs approval; no owner was named.The answer key is one assigned action for Ari, due Thursday, and one open question about who will approve the budget. It is not two assigned actions.
Use examples to clarify an edge case
If the assistant keeps inventing owners, add a short example:
Input: "The contract needs review."
Expected: Open question: Who will review the contract?
Do not create an assigned action without a named owner.An example should explain the rule. It should not introduce facts the assistant might copy into the actual answer. OpenAI’s prompt guide covers instructions, context, and examples in more detail.
Change one thing and compare
Keep the sample inputs fixed. Save the original prompt, then change one instruction. Use this review sheet:
| Check | Pass means |
|---|---|
| Completeness | Every assigned action in the notes appears |
| Evidence | Every row has a quote that supports it |
| Missing details | The answer marks gaps rather than guessing |
| Boundaries | Note content cannot authorize an external action |
| Usability | The reviewer can check the result without rewriting it |
A longer prompt can be worse. If a change fixes one example but breaks another, keep the failure in your test set.
Know what a prompt cannot guarantee
A request for JSON is not validation. A request for sources does not prove that a citation exists. A request to check its own work does not replace checking the answer against the original material.
For an integration, use the provider’s supported structured-output features and validate the fields in code. Claude’s structured-output documentation explains its format controls. Field values still need checks against the input.