Fix a vague AI answer with specific feedback

Replace “make it better” with a diagnosis, a constraint and a concrete revision request.

When a response is disappointing, a longer prompt is not always the answer. First identify the defect. Is the response generic, factually unsupported, badly structured or written for the wrong reader? Each problem needs a different correction. This method uses a small workshop announcement as an example.

Name the problem precisely

“This is bad” does not tell the model what to change. Try “The opening does not say what the workshop teaches” or “The draft invents a discount.” Quote the sentence that needs repair. Keep useful parts of the answer so the next revision does not discard them accidentally.

Supply the missing evidence

For a generic answer, add the facts that distinguish your situation: the topic, date, audience and practical benefit. For an invented claim, restrict the revision to the source notes. If the information is missing, ask for a question or an explicit gap rather than allowing the model to guess.

Try this revision prompt

EXAMPLE PROMPT

Revise the announcement below using only these facts: a beginner pottery workshop, Saturday 3 pm, 90 minutes, materials included, booking details to be added later. The current opening is vague and the draft invents a discount. Lead with what attendees will make. Remove the discount and unsupported urgency. Keep the friendly tone. Return a version under 100 words and list any facts still needed.

Change one major variable

If you simultaneously change the tone, length, audience and format, you will not know which correction helped. First fix facts and meaning. Then improve structure. Polish sentence style last. Keep a copy of the earlier version so you can compare rather than relying on memory.

Decide when to stop

Use a short acceptance checklist: correct facts, clear main point, suitable tone and required format. Once the draft meets those conditions, make the final human edits. Repeatedly asking for “more engaging” can add hype or remove useful detail. If the model continues inventing facts, reduce the task and work directly from the source material.