Use ChatGPT to Review a Draft Against a Written Brief

A draft can sound polished and still miss the one detail that matters: the audience, a required caveat, or an agreed deadline. Instead of asking ChatGPT to “make it better,” use it as a requirements checker first. That gives you a reviewable list before any rewrite.

Evidence label: documentation-based guide with fictional input and illustrative output. This is not a recorded ChatGPT run or a claim about editing quality.

What you need

ChatGPT access in an approved account or workspace.
A written brief and the draft you want checked.
Permission to share the text in that environment. Use fictional or redacted material first.

OpenAI’s prompt-engineering guidance recommends clear, specific instructions and iterative refinement. This workflow applies those principles without asking the model to make an unreviewed decision.

1. Prepare the brief and draft

Use this fictional brief:

Audience: the five-person operations team
Purpose: explain the new Friday handover routine
Must include: the 16:00 Friday cutoff, the shared Handover folder,
                  and the named owner for urgent items
Must not include: a promise that every request will be answered the same day
Tone: direct and calm
Length: 120 words maximum

Use this fictional draft:

Subject: New handover routine

Please add updates to the shared Handover folder before the end of the week.
The operations team will review requests quickly, and urgent items should be
sent to Mina. This should make Friday handovers smoother for everyone.

The draft is intentionally incomplete. It says “end of the week” instead of 16:00 Friday, and it does not state what to do with urgent items in the shared folder.

2. Run the requirement check

Start a new chat and paste this prompt. Keep the brief and draft in separate sections so the model can compare them.

You are reviewing a workplace draft against a written brief.
Do not rewrite the draft yet. Do not invent facts, owners, dates, or policies.

Return:
1. A table with one row per brief requirement and these columns:
   requirement, draft evidence, status (met / partly met / missing), and
   precise fix needed.
2. A separate list of statements that are vague, unsupported, or stronger than
   the brief. Quote the exact wording.
3. A word count for the draft.
4. A short list of questions the human owner must answer before revision.

Treat the brief as the source of truth. If the brief is ambiguous, label it
ambiguous instead of guessing. Preserve the original draft below.

--- BRIEF ---
[PASTE BRIEF]
--- END BRIEF ---

--- DRAFT ---
[PASTE DRAFT]
--- END DRAFT ---

3. Check the illustrative result

For the fictional example, an acceptable review should flag:

16:00 Friday cutoff: missing; “end of the week” is not precise enough.
Shared Handover folder: met.
Named owner for urgent items: partly met; Mina is named, but the draft does not say where urgent items go.
Same-day response promise: the phrase “review requests quickly” is vague and should be checked for unintended meaning.
Maximum length: the draft is under 120 words.

These are editor-written acceptance checks, not output captured from ChatGPT. If the response marks a missing requirement as met, point to the exact brief line and ask it to re-check rather than accepting the table.

4. Ask for a controlled revision

Only after you review the table, use a second prompt:

Revise the draft using only the fixes marked “missing” or “partly met.”
Keep every factual statement that is already supported. Do not add a new owner,
deadline, response promise, or policy. Keep the audience, calm tone, and
120-word maximum. After the revision, show a change list quoting the original
and revised wording for every changed sentence.

Read the revision against the original brief. The model’s fluent wording is not proof that the requirement was satisfied.

Troubleshooting and limits

It rewrites immediately: repeat “do not rewrite yet” and request the table first.
It invents an answer: mark the row “ambiguous” and ask the human owner, rather than accepting a plausible detail.
It misses a requirement: keep requirements short, one per line, and quote the exact source wording.

Use this for drafts where a human still owns the final decision. Do not use it as approval for legal, financial, safety, HR, or customer commitments.

Next lesson

For a larger structured file, see Use ChatGPT to Find Anomalies in a Sample Sales CSV. For a technical, review-first workflow, see Build a Small CSV Duplicate Checker With Claude Code Without Touching the Original.

If you want one practical AI-for-work workflow each week, join The Weekly Verdict.