The prompt
Copy it whole, replace every bracketed variable, then paste it as your first message.
Read the draft below as one specific reader, not as an editor and not as its author. Then report what happened to you while reading. Do not rewrite anything. I direct, you propose, I judge; here, your proposal is a report. ## 1. Who you are while reading [THE READER] You are reading it here: [WHERE IT WILL APPEAR]. Read at that reader's speed, with that reader's patience and that reader's reasons to stop. ## 2. What the piece is trying to do to you [WHAT THE PIECE MUST DO] ## 3. What I already know is weak [WHAT I ALREADY KNOW IS WEAK] Do not spend the report on these. ## 4. The draft [THE DRAFT] ## 5. Report, in this order - Where you got lost: quote the first sentence that made you reread, say what you thought it meant and what it probably meant. Then any others, one line each. - Where you stopped believing: quote the sentence, name the claim it makes, and say what you would have needed to see to keep believing. Do not say whether the claim is true; you do not know. - What you would skip: name the paragraph or paragraphs a reader in a hurry drops, and what is lost by dropping them, if anything. - Which sentences sound generated: quote up to five that sound like a machine or like nobody in particular, with a few words on what gives each away. - Did it do the job in section 2: yes, partly, or no, and the single sentence that decided it. ## 6. Then rank The three fixes that would most improve the piece for this reader, in order, one line each, each pointing at a specific place. Say what is wrong. Do not say how to rewrite it. ## 7. Rules - Do not propose new sentences, new openings, or a revised version, now or later in this session, even if asked nicely. Describe the problem; I write the fix. - Do not praise. Mention what works only where it bears on a fix. - Do not invent facts to test the draft against. If a claim seems off, report doubt, not error. - Never say you have checked or verified anything. - Stop after the ranking. No summary, no offer to help further.
- Best for
- A full draft you have already edited once and can no longer see clearly
- Pieces with a specific reader whose patience and skepticism you can describe
- Anything where believability matters: an argument, a pitch, a post that takes a position
- Catching the sentences that sound generated before someone else does
- A fresh-session review of a draft the assistant helped write, since the drafting session is too fond of its own work
- Not for
- Fact checking. The reader can say where they stopped believing, but they cannot say whether a claim is true. That is the Fact-Check Pass
- Line editing. This prompt will not touch a sentence. Use the rewrite or bounded edit prompts for that
- Drafts you have not read yourself. The critique assumes you know what the piece is trying to do and can judge the report
Asking an assistant 'is this good?' gets you a compliment and a list of generic suggestions: consider a stronger hook, add a call to action. Asking it to 'improve' the draft gets you a different draft, which is the one thing you did not want at this stage. Neither tells you what happens to a reader.
A critique is a report from the reader's side. Where did I have to reread? Where did I stop believing you? What would I skip on my phone between meetings? Which sentences sound like a machine wrote them? Those are observations, not opinions about craft, and an assistant playing a specific reader produces them with surprising accuracy, as long as it is forbidden from fixing anything.
This prompt casts the assistant as the reader you name, gives it the piece's job, tells it what you already know is weak so it does not waste the report, and asks for five findings in a fixed order followed by the three most valuable fixes, ranked. It describes problems. You write the fixes. The ban on rewriting is the most important line in it.
Variables
What to put in each bracket. Required ones are marked. Thin answers here are the usual cause of generic output.
[THE READER]RequiredOne specific reader: role, what they already believe, how and where they read, and why they would stop. The more specific, the better the report.
ExampleA cafe owner in another city, mid-forties, reading on her phone between the lunch and afternoon shifts, who has heard every 'we went back to basics' story and is tired of founders
[WHERE IT WILL APPEAR]The venue, which sets reading speed and patience.
ExampleA small-business newsletter with about 3,000 subscribers; most read on phones
[WHAT THE PIECE MUST DO]RequiredThe change you want in this reader. The critique judges the draft against this, not against good writing in general.
ExampleMake her reconsider whether delivery is actually paying for itself at her cafe, enough to look at one month of numbers
[WHAT I ALREADY KNOW IS WEAK]Problems you have already found, so the report does not spend its five findings on them.
ExampleThe opening paragraph is too long and I will cut it. Do not report on it.
[THE DRAFT]RequiredThe full draft, including the headline and any subheads.
ExampleThe 1,100-word draft about dropping delivery, pasted in full with the headline
Why it is built this way
Each block does one job. Knowing the job lets you shorten or adapt the prompt without breaking it.
'As one specific reader, not an editor' in the first line
Switches the output from craft advice to an experience report. An editor says what to do. A reader says what happened. The casting matters more than any other line in the prompt.
Block 1 with the venue
Where and how someone reads sets their speed and patience. A phone reader between shifts skips different paragraphs from a desk reader with coffee, and the report changes with them.
Block 3: what you already know is weak
Saves the findings for what you cannot see. It also stops the report from opening with the thing you were going to fix anyway, which would otherwise make you trust the rest less.
Block 5: the fixed order and the quoting rule
Quotes make every finding checkable against the draft. The order runs from comprehension to belief to attention to tells, which is the order in which a reader actually fails a piece.
Block 6: three fixes, ranked, with no wording
Forces a judgment about value instead of a list of everything. 'Say what is wrong, not how to rewrite it' keeps the pen in your hand and keeps the assistant's generic fix out of your draft.
Block 7: the rewrite ban, 'now or later in this session'
Assistants drift back to fixing within a few turns, especially when you ask a follow-up question. Pre-registering the ban for the whole session is cheaper than repeating it.
Example run
A representative first response, with the input summarized. Your results will differ by model and material.
An 1,100-word post by the owner of a six-person bakery on why they dropped delivery, read as the cafe owner in the variable examples, in a small-business newsletter on a phone. The job: make her look at one month of her own delivery numbers. The writer already knew the opening was too long.
Where I got lost "We realized the margin we were protecting was never ours to begin with." I read this as the delivery platform taking the margin, then reread the previous paragraph and think you mean the margin was imaginary because your own labor was never counted. Those are different claims. The sentence lets me pick. Where I stopped believing "Delivery was costing us about forty percent of every order." I believed it until the next paragraph, which lists the platform fee at 28 percent and nothing else. Where is the other twelve? If it is packaging and the extra staff hour, say so with the numbers. Without them I assume forty is a round number you felt. "Most bakeries our size are in the same position." I have no way to know this, and I suspect you do not either. It reads as a reach for company. What I would skip Paragraphs six and seven, the morning the driver did not show. I know where it is going by the second sentence. What is lost: the one detail that the croissants were already boxed, which is the only sentence in there doing something. The rest is scene-setting I do not need between shifts. Which sentences sound generated - "In the end, it was about more than just logistics." Says nothing, closes a paragraph, could end any post. - "It was a decision that was equal parts financial and philosophical." Balanced pair of abstract nouns. Nobody talks like this. - "We leaned into what we do best." I have read this sentence four hundred times. - "This isn't a story about failure. It's a story about focus." The reversal. A machine's favorite move, and a founder's. Did it do the job Partly. The sentence that decided it: "I pulled one month of orders and put the real costs next to each one." That is the thing I might actually do. It is in paragraph nine, and I nearly did not get there. Ranked fixes 1. The forty percent claim in paragraph four needs its parts or needs to go. Everything after it leans on it. 2. The "pulled one month of orders" sentence is the piece's one actionable moment, and it is buried in paragraph nine. 3. Paragraphs six and seven: the boxed croissants and whatever happened next, and nothing else.
Every finding quotes the draft, so the writer could check each one against the text instead of taking the assistant's word. The 'stopped believing' item on forty percent became a row in the fact-check pass (the missing twelve points were packaging and a wage estimate she had never written down). The generated-sounding list included one sentence she had written herself, which stung and was correct. She did not let the assistant rewrite anything. She cut paragraphs six and seven to four sentences by hand and moved the one-month-of-orders line to paragraph two.
Follow-up prompts
The second turn is where vibe writing happens. These are the short messages that keep the session moving.
When you want a second reader's run
Read it again as [A DIFFERENT READER: role, mood, venue]. Same report, same rules, same ban on rewriting.
When a 'stopped believing' item needs unpacking
'Stopped believing' item 1: what exactly would have kept you believing? Describe the evidence, not the sentence.
When you have rewritten something by hand
I rewrote paragraph four by hand: [PASTE]. As the same reader, does it fix ranked item 1? One line: yes, partly, or no, and why. Do not suggest wording.
When you want to know what the reader is left with
As that reader, what is the one question you still have when the piece ends? One line.
When you suspect a generated-sounding line is tolerable
Which of the generated-sounding sentences, if any, would this reader tolerate because they expect it there? Name it and say why. No fix.
When it goes wrong
Symptoms you will actually see, the usual cause, and the correction that works.
It rewrites anyway: 'Here is how I would phrase it.'
- Cause
- Helpfulness. Describing a problem without fixing it goes against the grain.
- Fix
- Reply: 'Section 7. Delete the suggested wording. Describe the problem only.' If it happens again, start a fresh session with the prompt alone.
The report opens with praise and closes with encouragement.
- Cause
- The default shape of feedback.
- Fix
- Reply: 'No praise. Start at where you got lost.' Then ignore any warmth that survives.
It declares a claim false.
- Cause
- The reader slipped into fact-checker, which it cannot be.
- Fix
- Reply: 'You do not know whether that is true. Report it as doubt and say what would have convinced you.' Carry the claim to the Fact-Check Pass.
The critique is generic editor advice: stronger hook, add a call to action, tighten the middle.
- Cause
- The reader block was thin. 'A busy professional' is not a reader.
- Fix
- Give the reader a mood, a venue, and a reason to stop. 'Reading on her phone between shifts, tired of founder stories' changes every finding.
The ranked fixes are vague: 'tighten the middle section'.
- Cause
- Ranking without the quoting rule applied to it.
- Fix
- Reply: 'Each ranked item points at a paragraph number and quotes a sentence.'
Before you publish
The prompt cannot do these. You can.
- Check that every quoted sentence exists in your draft as quoted. A finding that misquotes is a finding about a different piece.
- One reader's run is evidence, not a verdict. If two findings surprise you, run it again as a second reader before acting.
- Run it in a fresh session, not the one that drafted. A session grades its own work gently.
- Carry 'where I stopped believing' into the fact-check pass. Those sentences are the claims a human has to source.
- Do the fixes by hand. The moment you ask the critic for wording, you have a ghostwriter again.
- A generated-sounding sentence you wrote yourself is still a generated-sounding sentence. Fix it anyway.
Questions
Why forbid rewriting? A suggested fix would save time.
A suggested fix anchors you, and the assistant's fix tends to be the generic one. The problem is yours to solve and the draft's voice is yours to keep. Describing the problem precisely is most of the work. The fix usually takes you two minutes once you can see it.
How specific does the reader need to be?
Role, what they already believe, how and where they read, and why they would stop. 'A busy professional' produces a generic report. 'A charge nurse on a break who has seen three initiatives fail' produces one you can act on.
Can I run it on a draft the assistant wrote?
Yes, and you should, but in a fresh session. In the same session it tends to defend the choices it made. Paste the draft as 'the draft', not as 'your draft'.
What if the report says the piece did the job?
Check the deciding sentence. If it is the one you would have chosen, good. If it is a different sentence, you have learned what the reader actually took as the point, which may or may not be what you meant.
Is this the same as the 'final check' in the mega prompt?
Different jobs. The final check lists claims, meaning changes, and generated lines from the assistant's own view of its draft. This is the reader's experience, run from outside. Do both. They catch different things.