TaskbyTask.Practical AI guides for work
ChatGPT, Claude, Gemini or Copilot · Practical walkthrough

Improve an answer without starting the whole conversation again

The draft is nearly useful. Asking for a complete rewrite may lose the good bit as well as the bad bit. Tell AI what to keep and what to change.

Task by Task13 min read + practice
A partly repaired teal mosaic with one new lime tile fitted among intact ceramic pieces.

What you'll make

a revised draft with one checked improvement, the facts unchanged, and a short list of what the AI changed.

What you'll need

the draft in your current chat in ChatGPT, Claude, Gemini or Copilot, the facts it was built from, and one clear reason for the change. Allow about ten minutes.

Example result.

Before: “We have been considering ways to make our triage process more consistent and would welcome thoughts on the possibility of exploring a checklist trial.” After: “Please decide whether we should run a two-week trial of the triage checklist. The trial is still proposed, and we need two volunteers before it can begin.” Changes: 1. Moved the decision request to the first sentence. 2. Kept “proposed”. 3. No facts added.

1. Name the one thing to change

Read the draft and find the weakest part. “Too formal” might mean long sentences, an indirect opening or words the reader wouldn’t use. Name the visible problem and what the reader needs instead.

Pick one change. If you ask for a warmer tone, a shorter length, a new structure and a different conclusion all at once, you won’t know which instruction caused a useful change, or a harmful one.

Why this works: one named change gives you one thing to check. Everything else in the draft should come back the same.

What you should see: a single sentence you could hand to a colleague, such as “The opening hides the decision we need.”

2. Send a revision brief in the same chat

Most of the time, send the prompt below as your next message. Some tools also let you change one passage in place or try another direction without losing the current version:

Tool Where to make the change
ChatGPT If the draft is in a writing block, select the passage and ask ChatGPT to revise just that selection. Undo reverses recent AI edits
Claude In a document artifact, highlight the passage, click Edit with Claude and type the change. Editing an earlier message creates a separate version of the chat, so the current one is kept
Gemini Send the brief as a follow-up. Regenerate works on the most recent response only; the arrows under it switch between versions
Copilot Send the brief as a follow-up in the same chat

Regenerating is worth trying when the whole answer missed but your request was right. When most of the draft is good, revise instead: a fresh roll of the dice can lose the good bit too.

Revise the draft below for [reader and purpose].

Keep: [accurate facts, useful sections and wording that should remain].
Change: [one specific weakness].
Target: [a concrete result, such as putting the decision request first].
Do not add new facts, promises, dates or owners. Keep qualifications and unresolved information visible. Return the revised draft and a short list of the substantive changes you made.

Source facts:
[source facts and draft]

[source facts and draft] is the facts the draft was built from, then the current draft, both pasted in full. The other three brackets are your step 1 sentence, split into keep, change and target.

Why this works: the keep list protects what’s already good, and the target tells the AI what “better” looks like. Pasting the facts again lets you check whether a change of style has quietly changed the meaning.

What you should see: the revised draft, followed by a numbered list of changes you can check one by one.

3. Compare the revision with the facts

Check the improvement you asked for first. If you wanted the decision request at the top, is it there? Could the intended reader tell what happens next?

Then check what shouldn’t have moved: dates, figures, names, conditions and anything still unresolved. Use the change list to find the edits, but read the text itself. The AI may leave an important change out of its own summary.

Why this works: a smoother sentence is the most common place for meaning to shift, because the AI is optimising for how it reads.

What you should see: every fact from your list still present, with the same strength.

Check the result

  • The change you asked for is visible in the draft.
  • Each date, figure, name and condition matches the facts you pasted.
  • The commitment verbs haven’t moved: “could” is still “could”, “proposed” is still there, and “subject to approval” hasn’t been removed to make a sentence smoother.
  • The change list matches what actually changed in the text.

The usual failure: a stronger opening becomes a stronger promise. The draft now reads well, but a proposal has turned into an announcement, and the change list doesn’t mention it. Spot it by finding the verb that carries the commitment and comparing it with the facts.

Keep the revised structure. Restore this qualification from the source: [exact qualification]. The draft currently says [incorrect wording], which changes its meaning. Replace that passage only, keep the other checked facts intact and show the corrected passage in context.

[exact qualification] is the wording from your facts; [incorrect wording] is the sentence that overstates it. Then read the whole affected paragraph, not only the new sentence: another line may still imply the same commitment. Stop when the draft meets the purpose and the facts. Another round is only useful when you can name what it will improve.

Worked example: a trial request that needs a clearer opening

Fictional example. A service team wants approval to trial a new triage checklist. These are the facts it pastes into the prompt:

- Proposal: a two-week trial of a new triage checklist.
- Status: proposed, not approved.
- Existing response commitments stay unchanged.
- Two volunteers are needed; neither has been selected.

The draft opened with “We have been considering ways to make our triage process more consistent and would welcome thoughts on the possibility of exploring a checklist trial.” The team asked to keep the facts, change the indirect opening and put the decision request first.

The first revision came back as: “We are launching the new triage checklist for a two-week trial. Existing response commitments will remain unchanged, and we will confirm two volunteers this week.” Its change list said only “Made the opening more direct.” That’s the usual failure: “launching” announces a decision nobody has made, and “will confirm this week” invents a date.

After the repair prompt, with “The trial is still proposed, not approved” as the qualification and “We are launching” as the incorrect wording, the opening read: “Please decide whether we should run a two-week trial of the triage checklist. The trial is still proposed, existing response commitments would remain unchanged, and we need two volunteers before it can begin.” The change list now read: 1. Moved the decision request to the first sentence. 2. Restored “proposed”. 3. Removed “this week”, which wasn’t in the facts.

  • The verb carried the risk. “Considering” to “please decide” is a better opening. “Considering” to “launching” is a different decision.
  • The change list missed the important edit. That’s why step 3 checks the text, not only the summary.
  • The repair touched one passage. The new structure survived; only the overstated sentence changed.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Fix the order before polishing sentences

If a draft buries its purpose, fix the order first. Polishing each sentence can produce elegant paragraphs that still arrive in the wrong sequence.

Ask for an outline with the reader’s question at the top. Agree the structure, then ask for the text, with the facts pasted alongside throughout.

Before rewriting, give me a three-to-five-line outline of this draft for [reader], with the question they most need answered at the top. Don't write the full text yet.

Make it shorter without losing the condition

Length is the change where qualifications disappear fastest. Try it on the worked example: the same request is accurate but has to fit a two-line Teams message. Before you prompt, write the Keep, Change and Target lines yourself. Keep: still proposed; two volunteers needed. Change: length. Target: two lines that end with the question.

If the short version drops “proposed” to save words, it’s too short. For longer documents, write a short list of facts that must survive every version and check each draft against it; the Google Docs rewrite guide shows that method inside a document.

Know when to stop revising

If each revision fixes one problem and introduces two more, the chat has probably collected conflicting instructions. Know when to start a fresh chat covers that decision and what to carry across.

Sources

Optional analytics

With your permission, Google Analytics uses cookies to measure visits and which guides people read. It stays off until you accept. Use Analytics choices in the footer to change your choice. Rejecting after accepting refreshes this page. Privacy details.