ChatGPT can also comment on structure, clarity, tone, and sentence-level errors. Choose the level of feedback you need. Correct grammar will not repair a recommendation whose evidence does not support its conclusion.
Tell the reader's task to the reviewer
“Give feedback on my writing” leaves the reviewer to invent the purpose. A useful brief says, for example, “This note asks a facilities manager to approve a small trial; it does not ask for a permanent rollout.” Include the length constraint and any required criteria.
OpenAI's prompting guidance, checked October 24, 2026, recommends adding context and describing the result and boundaries. Here, the boundary is keeping the author's ideas and supplied evidence intact while identifying what needs work.
For school assignments, include the actual assignment and permitted use of AI. For workplace writing, supply only material you are allowed to share in the approved account. Neither a polished rewrite nor a positive critique establishes that the draft meets an unstated requirement.
Ask for a diagnosis you can act on
Give developmental feedback on my draft. Do not rewrite it yet.
Audience: [WHO WILL READ IT]
Purpose: [WHAT THE READER SHOULD UNDERSTAND OR DECIDE]
Required criteria and length: [PASTE]
Available evidence: [PASTE OR IDENTIFY THE SUPPLIED SOURCES]
Identify the three most consequential revision priorities, or fewer
if fewer are needed. For each, quote the relevant draft wording,
explain the effect on this reader, and propose one concrete revision
action. Separate missing evidence from unclear explanation.
Do not invent facts, quotations, results, or the author's intentions.
Ask about an ambiguity rather than silently resolving it. Leave
grammar and optional word preferences until after the main issues.
Draft: [PASTE]
The quotation anchors a comment to something you can inspect. “Improve flow” gives you less direction than “Move the requested trial before the background, because the manager does not learn what approval is needed until paragraph four.”
Distinguish an evidence problem from a wording problem
Consider this original fictional memo fragment:
Three volunteers liked the new sign at the equipment desk. We should replace every sign this month because the new design eliminates collection mistakes.
The problem is larger than the word “liked.” Three volunteers' preferences do not establish that mistakes were eliminated. A useful critique would ask what the volunteers were asked, whether collection mistakes were measured, and why replacement across all desks follows from this small observation.
A narrower revision could be:
Three volunteers preferred the new equipment-desk sign. I propose testing it at one desk and recording collection mistakes before deciding whether to replace the other signs.
That is an authored example of a possible revision, not a real trial or a ChatGPT transcript. The proposed test still needs an owner and a workable measure. The revision does not turn an unmeasured outcome into an observed success.
If the feedback merely changes “eliminates” to “dramatically reduces,” it preserves the same unsupported claim. Reject that suggestion and ask the reviewer to separate what the evidence shows from what the author proposes doing next.
Apply one revision priority and read again
Decide which comments match the actual purpose. Feedback can be wrong about your audience or can demand detail that belongs elsewhere. You can reply, “The manager already has the cost sheet; focus on whether this note makes the trial decision clear.”
Revise the argument or order yourself, or request a bounded revision that names the accepted changes. Then compare the revised text with the original evidence. OpenAI's review guidance calls for checking consequential claims and assumptions before using an answer.
Once the reasoning and structure work, use a minimal proofread for spelling, grammar, and punctuation. If the draft depends on external claims, check their citations as a separate step. Finish by reading the opening and requested action together: the reader should know what you want and what the supplied evidence actually supports.