Get feedback on your writing before you send it
Why "can you improve this?" gets you a rewrite you cannot judge, and the four questions that get real feedback on writing you have already finished.
After five versions of a six-hundred-word announcement, Nadia can no longer tell how it reads. Nadia is telling the members of her pottery studio that Saturday open-studio ends in November and a booking system takes its place. The fifth draft is on screen. She knows the words are there. She has no idea any more what they do.
Her partner read it and said it was fine. That is what partners say at half past nine on a Sunday, and she knew it was coming before she asked. Everyone else she could send it to is a member of the studio, which makes them the audience rather than a test audience.
This is the ordinary situation, and it has almost nothing to do with grammar. Nadia does not need a comma checked. She needs somebody to tell her whether a member who reads this on a phone in about ten seconds will come away thinking the studio is closing.
The question that ruins it
The first instinct is to paste the notice into a chat and ask: can you improve this?
Something comes back within seconds. It will be six hundred words, it will be clean, and it will be different in ways you cannot assess. "Improve" carries no target, so the model supplies one of its own: shorter sentences, a warmer opening, headings where you had paragraphs, a closing line thanking everyone for their understanding. That is a house style being applied to your notice, and it gets applied whether or not your notice had a problem. You now hold two versions of the same document and no way to choose between them, because you asked for prose and prose is what came back.
Rewrites are the easiest thing in the world for a model to produce and the hardest thing for you to judge. Feedback is the other way round. So ask for feedback, and ask for it one question at a time.
Ask four smaller questions instead
Nadia starts a fresh chat, pastes the notice, and spends one sentence on who is reading it: around ninety members, half of them come in once a month, most will open this on a phone within a minute of the notification. Feedback without an audience is decoration.
Whatever you paste is unpublished, so it is worth knowing where it goes. We do not train on your conversations, and what the provider behind a given model does varies; only six of the 60 say so on their info card.
What will a reader misunderstand? This is the highest-value question on the list, because it is the only failure that costs you anything after you press send. A useful answer is specific about where and why:
Paragraph two says "the Saturday session as it currently runs will end." A reader skimming will stop at "will end" and conclude that Saturday opening is finishing. The replacement booking system is not mentioned until paragraph four, which is below the fold on a phone. Anyone who reads only the first screen will think the studio is dropping Saturdays.
That is a real note about a real reader, and she can check it herself in five seconds by reading the first screen the way a member would.
Which paragraph is weakest, and why is it weak? Asking for a ranking forces a choice, and the "why" is the half you use. "The third paragraph is weakest because it explains the reason for the change twice, once in cost terms and once in fairness terms, and the two reasons undercut each other" tells you what to do. "The third paragraph could be tightened" does not.
Where do I lose the thread? Phrased that way it asks for a location, the point where a reader stops following, and that is rarely the sentence you were worried about. In Nadia's case it was a subordinate clause about the studio's insurance renewal, which she had put in because it was true and which no member needed.
What am I assuming the reader knows? The one that catches your own expertise looking back at you. She uses "kiln slot" four times. Roughly a third of her members have never booked one and do not know that it is separate from a bench.
Then a fifth, once the first four have been answered: what is already working, so I do not cut it by accident. Ask it last. Asked first, it sets a tone of approval that the rest of the reply then has to fight.
Praise arrives first, and it is worth nothing
Nearly every model opens with a compliment. Warm, clear, well structured. It is reflexive rather than dishonest, the conversational throat-clearing that comes before the content, and the safest thing to do with it is skip to the second paragraph.
There is a simple test for the rest. Feedback that names a location is feedback: paragraph two, second sentence, the clause about insurance. Feedback that describes an atmosphere is weather. "The tone is welcoming and professional" survives being pasted onto any document ever written, which is what tells you it was not really about yours.
One more thing it will not volunteer: that the piece should be half as long, or that it should not be sent at all. Models answer the question in front of them. If you want to know whether the whole notice is a mistake, that has to be its own question.
Why your voice goes missing here
Here is the part that matters most, and where most people quietly lose their writing.
Accept every suggestion in the reply and you end up with something competent that nobody wrote. Every edge is sanded, every joke moved to a safer position, every specific detail traded for a general one. For a studio whose whole appeal is that a person runs it, publishing that under your name is a real loss, and the members will not be able to say why the notice felt off.
The irony is worth stating plainly. The constructions a model reaches for while smoothing your prose are the constructions readers have learned to identify as machine-written. The three-item list where two items would do. The sentence that denies one thing in order to affirm another. "Seamless", "delve", "at the end of the day", "we are excited to announce". Handing your writing over to be tidied is a fast way to make it read as though no person was involved.
So split the reply in two. When it says paragraph two misleads a skimming reader, that is information about your reader, and it is the reason you asked. When it hands you a replacement paragraph two, that is a sample of its voice, and you already have one of those. Take the diagnosis and write the fix yourself, in your own words, badly if necessary.
Nadia's replacement opens: "Saturdays are not going anywhere. From November you book your bench instead of turning up and hoping." Shorter than what the model offered her, less balanced, and unmistakably hers.
A second reader catches what the first one missed
One reader's blind spot is not another's, and this holds for models as much as for people. Nadia pastes the same notice, with the same four questions, to a different model. It says nothing about paragraph two. It asks what happens to the members who have already paid for a Saturday block through to December, which the notice never answers, and which the first model read straight past.
The rule of thumb is easy: a note about your text that two models make independently is almost certainly real, and a note only one makes is worth ten seconds of your own judgement. What you are doing is closer to circulating a draft than to running a checker.
Keeping both answers takes one deliberate choice, and the menu labels work against you here. Regenerating replaces an answer: the old one is deleted and there is no undo, so it is the wrong control. Branching is the one you want. It opens a fresh thread that carries the conversation up to that point and leaves the original where it was. On the web the branch action is labelled Try Again. In the mobile app it is Branch here, and Try again on a phone means regenerate. Check the label on the device you are actually holding. Two threads, two readings, and you switch between them.
Where to keep a standing instruction about your voice
By the fourth time she does this, Nadia is retyping the same preamble: keep my voice, no bullet points, British spelling, do not rewrite unless I ask.
That belongs in Personal preferences, in your settings: a plain text box attached to every new conversation. It needs an account and nothing else. Saved memories are on the paid plans; this box is not. Something like "When I ask for feedback on my writing, point at problems and do not supply replacement text unless I ask for it" is a single sentence that changes every review you run from then on.
Ten minutes of reading, then close the chat
A model is a serviceable first reader and a poor second author. The best use of it is the ten minutes after you have finished and before you send, when your own eyes have gone and there is nobody around who will risk telling you the truth.
Nadia's sixth version has an untidy sentence in it that no editor would have left. It is also the one that stopped ninety people thinking her studio was closing.