Give AI Feedback on One Failure Instead of Rewriting the Whole Prompt

When an assistant produces the wrong result, adding several new instructions at once can make the next attempt harder to understand. Identify one observable failure and revise the instruction that governs it. Keep a record of the original output.

Describe the failure concretely

Replace “The answer is bad” with “The table omitted two tasks whose status field was blank.” This identifies the affected records and the behaviour you expected. Before blaming the assistant, confirm that your instructions explained how blanks should be handled.

In a fictional task list, define blank status as Unknown rather than Complete. Add that rule to the prompt and ask for the same extraction again. Avoid changing the format and the scoring rules in the same attempt, because you would no longer know which change helped.

Retain a small regression set

Keep one ordinary record, one blank field and one unusual value as test inputs. Write the expected output for each. After a revision, check all three; fixing the blank case should not silently break a previously correct normal case.

Ask the assistant to explain any unresolved uncertainty, but do not accept its claim that all tests passed without comparing the output. A short checklist maintained by you is more useful than an unsupported confidence score.

Decide when to stop iterating

If the task keeps failing, reduce its scope or use a more deterministic method for the fragile part. A spreadsheet filter may be easier to verify than repeated natural-language extraction for a simple counting job.

Save the final prompt, sample input and observed result together. Describe the improvement narrowly: it handled these examples better. Do not turn a successful small test into a claim that the prompt is reliable for every future document.