# Fix a weak AI answer without starting again

Original 5HO practice material. Fictional example. Reviewed 5 October 2026.

## Useful result

You'll finish with a revised answer with the actual weakness addressed.

When an answer misses the mark, change the instruction that caused the miss. “Try harder” gives the model very little to work with. A precise correction gives you something you can test.

## Set up

Use 10 minutes, the draft below and any AI chat. Keep the original answer so you can compare revisions instead of trusting the newest one automatically.

## Practice input

Brief: write a 60-word factual update. Draft: Our exciting rollout is complete and everyone loves it. Evidence: pilot started Monday; 4 of 6 users tried it; two reported login problems; feedback review Friday.

## Do the task

1. Identify the main fault: unsupported facts, unsuitable tone, wrong length or missing structure.
2. Request one correction using the evidence, then check whether new errors appeared.
3. Stop after two unhelpful revisions and rewrite the brief or change the task.

## Prompt

Repair this draft using the supplied evidence. Remove claims of completion and universal approval. Keep it under 60 words and include the login issue and Friday review.

## Refine

Make it calm and direct without hiding the problem. Count the words, then show any remaining uncertain claim separately.

## Expected checks

The revision should report a pilot in progress and two login reports. Four trial users does not mean four satisfied users. Repeatedly incorrect facts are a reason to stop and inspect the input.

## Save

Keep the corrected draft plus a note of the instruction that fixed it. Reuse the correction when the same problem appears.

## Sources

- Writing clear prompts: https://help.openai.com/en/articles/10032626-prompt-engineering-best-practices-for-chatgpt