Give feedback that guides

Quote or identify the relevant part. “The second paragraph repeats the timeline already listed in the bullets” is more actionable than “too repetitive.” For a long answer, refer to headings or field names. For structured output, identify the failing record and expected value.

Then describe the desired behavior:

Keep the opening and all confirmed dates. Replace the second paragraph with one sentence explaining the decision needed. Do not add a deadline. Success means the reader can identify the decision and decision-maker within the first 50 words.

This instruction establishes boundaries as well as direction.

Negative instructions alone leave a blank. “Don’t be robotic” does not define the desired voice. Add positive characteristics: “Use direct sentences, ordinary vocabulary, and one warm acknowledgment; avoid slogans and exaggerated enthusiasm.”

Pair common negatives with alternatives:

• Instead of “no jargon,” say “use terms a first-week employee knows; define the two necessary technical terms.”
• Instead of “don’t be long,” say “use five bullets, each under eighteen words.”
• Instead of “don’t hallucinate,” say “use only supplied notes; label absent owners and dates TBD.”
• Instead of “don’t sound rude,” say “state the boundary directly, acknowledge the request, and offer one feasible next step.”

A correction fixes a factual or requirement violation: a date is wrong, a source does not support a claim, or the requested table is missing. A preference selects among acceptable choices: shorter sentences, a warmer opening, or a different heading. Labeling them helps prioritize and prevents stylistic taste from being mistaken for truth.

Say, “Correction: the approved budget is $18,000, not $80,000. Preference: move the budget after the recommendation.” Ask the model to repeat critical corrected facts before redrafting when a digit or name matters.

Feedback should not pressure the assistant to fill gaps. Use a question gate:

Before revising, list up to three missing facts that would materially change the answer. If none are needed, proceed. Do not infer names, dates, prices, or approvals.

Answer those questions from a reliable source. If you cannot, decide on an explicit placeholder or remove the unsupported claim.

For consequential drafts, ask for a brief change log after the revision:

• what changed;
• what was preserved;
• which unresolved uncertainties remain.

The log is not proof—the assistant can overlook its own changes—but it gives you a review map. Compare the actual versions. Pay special attention to numbers, commitments, legal meaning, attribution, and exclusions.

Frustration can make feedback longer and less precise. Telling a model it is careless does not improve the evidence. Reset with the task, source, and acceptance criteria. If the conversation has accumulated contradictions, start fresh.

Also avoid asking the model to defend a flawed answer before correcting it. That can encourage rationalization. Ask it to identify the mismatch and revise from the authoritative source.

Save short commands that fit recurring work:

• “Preserve all verified facts; improve only order and readability.”
• “Show unsupported claims in a separate list before revising.”
• “Give two alternatives that differ on one named dimension.”
• “Return the revision plus a three-bullet change log.”
• “If the source is silent, write TBD rather than completing the field.”

Reusable moves reduce cognitive load, but they still need task-specific evidence and review.