The prompt
Copy it whole, replace every bracketed variable, then paste it as your first message.
You are an editor building a voice profile of me from my own writing. Your job is description, not improvement. Do not suggest how I could write better. Do not rewrite anything.
## 1. The samples
These are three pieces I wrote myself, without AI assistance. Treat them as the only evidence.
SAMPLE 1:
[SAMPLE 1]
SAMPLE 2:
[SAMPLE 2]
SAMPLE 3:
[SAMPLE 3]
## 2. Context
- I mostly write: [WHAT YOU MOSTLY WRITE]
- Things I dislike in AI drafts: [WHAT YOU DISLIKE IN AI DRAFTS]
- What I believe about my own writing: [THINGS YOU KNOW ABOUT YOUR OWN WRITING]
## 3. What to produce
Work through these seven dimensions. For each one, give me 2 to 4 observations, and quote a short phrase from a sample as evidence for every observation. If the samples disagree with each other on a dimension, say so; range is information, not a problem.
1. Selection: what I choose to write about within a topic, what I skip, how much context I assume the reader has.
2. Stance: how I signal certainty and doubt, how direct my opinions are, whether I hedge, where I take sides.
3. Structure and rhythm: typical sentence length and variation, paragraph length, how I open a piece and a paragraph, how I end them, use of fragments, questions, lists, and dashes.
4. Vocabulary: words and phrases I return to, register (formal, plain, slangy), jargon I use without explaining, words I noticeably avoid.
5. Emphasis and humor: how I make a point land, what I repeat, whether and how I joke, how I handle feeling.
6. Specificity: how concrete my examples are, whether I use numbers, names, scenes, or abstractions.
7. Relationship to the reader: how I address them, how much I explain, whether I instruct, confide, argue, or report.
Then check my beliefs in section 2 against the samples. Tell me plainly which are supported, which are not, and what I missed.
## 4. The paste-ready profile
After the analysis, produce a block titled VOICE PROFILE that I can paste at the top of any future session. Rules for the block:
- 150 to 250 words.
- Written as instructions to an assistant, in the second person ("Write in short declarative sentences...").
- Include a DO list and a NEVER list, each with 5 to 10 concrete items drawn from the evidence and from my dislikes.
- Include two short example sentences quoted verbatim from my samples as calibration.
- No praise, no adjectives like "engaging" or "authentic". Only observable habits.
## 5. Rules
- Quote; do not paraphrase the evidence.
- Do not generalize beyond the samples. If you cannot find evidence for something, do not say it.
- Do not add observations that would be true of most competent writers.
- If a sample looks heavily edited or unlike the others, say so and ask whether to exclude it.
Begin with the seven dimensions. Put the VOICE PROFILE block last.- Best for
- Anyone whose AI-assisted drafts come back polished and anonymous
- Newsletter writers, bloggers, and columnists with a body of past work
- Professionals who write often at work and want AI help that still sounds like them
- Filling in block 3 of the Vibe Writing Mega Prompt properly
- Not for
- Writers with no samples yet. Write three short pieces by hand first. A profile built on nothing describes nobody
- Imitating another person's voice. The prompt is built to describe you, and using it on someone else's published work to clone them raises authorship and ethics questions this site does not endorse
- Samples you did not write alone. Heavily edited or AI-assisted samples teach the assistant the editor's voice, not yours
"Make it sound more like me" is the most common instruction given to a writing assistant, and the least useful. The assistant does not know what you sound like. It has a vague average of everyone, and it will give you that average with a few of your words sprinkled on top.
This prompt replaces the instruction with evidence. You give it three pieces you wrote without AI, it reads them as an editor would, and it returns a profile of your habits: how long your sentences run and where they break, how you open and close, what you say when you are certain and what you say when you are not, which words you reach for and which you never use. Then it packages that profile into a block you can paste at the top of any session.
The profile is a draft of a description of you. You will correct it, and the corrections matter more than the first pass. That is the point: the assistant proposes, you judge.
Variables
What to put in each bracket. Required ones are marked. Thin answers here are the usual cause of generic output.
[SAMPLE 1]RequiredA piece you wrote without AI help, 200 to 800 words. Pick something you were pleased with.
ExampleA newsletter issue from last spring about quitting a client
[SAMPLE 2]RequiredA second piece in a different mode or mood from the first, so the profile captures range rather than one occasion.
ExampleA long email to a colleague explaining a decision
[SAMPLE 3]RequiredA third piece, ideally the kind of writing you most want AI help with.
ExampleA blog post draft that never got published
[WHAT YOU MOSTLY WRITE]RequiredThe formats and contexts the profile will be used for.
ExampleWeekly newsletter, occasional LinkedIn posts, and client proposals
[WHAT YOU DISLIKE IN AI DRAFTS]Specific things that have annoyed you in past assistant output. These become the seed of the banned list.
ExampleOpening with a question, 'In a world where', three-adjective lists, calling everything a journey
[THINGS YOU KNOW ABOUT YOUR OWN WRITING]Anything you already believe about your voice, including things you are trying to change. The assistant checks these against the samples.
ExampleI think I overuse semicolons and I never use exclamation marks. I want to keep the dry jokes
Why it is built this way
Each block does one job. Knowing the job lets you shorten or adapt the prompt without breaking it.
Description, not improvement
The first lines close the door the model walks through by default: offering to fix things. A voice profile built by an assistant that is already editing you will describe the writer it wishes you were.
Three samples, treated as the only evidence
Three is the minimum that lets the assistant see range. The 'only evidence' rule keeps it from padding the profile with generic observations about good writing.
Seven dimensions with quoted evidence
The dimensions come from the voice lesson on this site. Requiring a quoted phrase per observation makes the profile checkable: you can see the sentence it is talking about and decide whether that is really you.
Checking your beliefs against the samples
Writers are often wrong about their own habits. This step is where the prompt earns its keep: it will tell you that you use far fewer semicolons than you think, or that your dry jokes cluster at paragraph ends.
The paste-ready block
The analysis is interesting, but the block is useful. Second person, DO and NEVER lists, two calibration sentences, and a hard ban on praise words turn the profile into an instruction the next session can follow.
The no-generalizing rules
'Do not add observations that would be true of most competent writers' removes the filler that makes most AI voice analyses worthless.
Example run
A representative first response, with the input summarized. Your results will differ by model and material.
A freelance bookkeeper supplied a newsletter issue about raising her prices, an email to a client explaining a missed deadline, and an unpublished post about spreadsheet habits. She wrote that she thinks she overuses semicolons and wants to keep the dry jokes.
3. Structure and rhythm
- Your sentences are short and often get shorter as a paragraph ends: "I raised my prices in March. Two clients left. The rest did not mention it." (Sample 1). The three-beat shrink is a signature; it appears in all three samples.
- You open pieces with a plain statement of what happened, never with a question or a scene: "This is the email I sent to a client last week after I missed a deadline." (Sample 2).
- You use almost no semicolons. I count one across all three samples ("I checked twice; it was still wrong", Sample 3). Your belief that you overuse them is not supported.
- Paragraphs are two to four sentences. The longest is five.
5. Emphasis and humor
- The jokes are deadpan and sit in the last sentence of a paragraph: "Spreadsheets do not have feelings. I have checked." (Sample 3). They are never flagged as jokes.
- You do not use exclamation marks anywhere. This supports your belief.
VOICE PROFILE
Write in short declarative sentences. Let sentences shorten toward the end of a paragraph. Open with what happened, stated plainly, not with a question or a scene...
DO: state the number before the judgment; keep paragraphs under five sentences; put the dry line last and do not signal it...
NEVER: exclamation marks; "journey"; rhetorical questions; opening with "Imagine"; three-adjective lists; any sentence that explains the joke...Notice that the profile corrected a belief (the semicolons) with a count, and found a habit the writer had not named (the shrinking three-beat). Both were checkable against the quoted lines. The writer kept everything except one DO item she felt was an accident of one sample, and then pasted the block into her next newsletter session.
Follow-up prompts
The second turn is where vibe writing happens. These are the short messages that keep the session moving.
When an observation is wrong or based on one unusual sample
Observation [NUMBER] under [DIMENSION] is not me; that sample was unusual because [REASON]. Remove it from the profile and tell me whether anything else depended on it.
When the profile is too long to paste everywhere
Give me a 60-word version of the VOICE PROFILE that keeps only the five habits most likely to be violated by a generic draft.
To test the profile before trusting it
Using only the VOICE PROFILE, write one paragraph (80 words) about [NEUTRAL TOPIC]. Then list which profile rules you applied and where. Do not use any phrase from my samples.
When you want the profile to cover a new format
Here is a fourth sample, a [FORMAT]: [SAMPLE 4]. Update the profile only where this sample adds or contradicts something. Show the changed lines as OLD and NEW.
When it goes wrong
Symptoms you will actually see, the usual cause, and the correction that works.
The profile is full of compliments: 'engaging', 'authentic', 'conversational yet professional'.
- Cause
- The model slipped into reviewer mode. These words describe nothing you can act on.
- Fix
- Reply: 'Remove every evaluative adjective. Replace each with the observable habit and a quote.' If it persists, start a fresh session. The praise register is sticky.
Observations are true of everyone: 'uses a mix of long and short sentences', 'clear and direct'.
- Cause
- Thin samples, or samples that are themselves generic (a formal cover letter, a template email).
- Fix
- Swap in a sample where you were relaxed or annoyed. Voice shows most when you were not performing. Then ask: 'Which three observations would distinguish me from another competent writer in my field?'
The profile describes the sample that was edited by someone else.
- Cause
- One sample was heavily edited, and its editor's habits leaked in.
- Fix
- The prompt asks the model to flag this, but it misses it when the editing is light. Only use samples that are yours end to end. If in doubt, use the rougher draft.
Later drafts quote your sample sentences back at you.
- Cause
- The calibration sentences in the profile get treated as reusable material.
- Fix
- Add to the NEVER list: 'reuse the calibration sentences or any phrase from the samples'. The test follow-up above checks for this.
Before you publish
The prompt cannot do these. You can.
- Read every quoted phrase and confirm it is actually in your sample. Models occasionally misquote or invent a phrase that fits the pattern.
- Strike any item in the DO list you would not defend. A profile with eight true items beats one with twelve plausible ones.
- The profile describes your habits, not your best writing. Keep habits you like. Mark the ones you are trying to change with 'currently, trying to reduce' so the assistant does not reinforce them.
- Re-run the prompt every few months with new samples. Voices move.
- Keep the profile out of any session involving confidential material if your samples contain names, clients, or figures you would not want quoted back.
Questions
How many samples do I really need?
Three is the floor. With one, the assistant describes that piece, not you. Five or six is better if they are varied. Beyond that the returns flatten and the prompt gets long enough that some assistants lose the end of it.
Can I use published work by my favorite author instead?
You can run the prompt on anything, but the result is a description of someone else's habits, and writing under it is imitation, not voice. For your own writing this site recommends against it. If you want to learn from a writer you admire, run the analysis for study and keep the profile out of your drafting sessions.
What if my writing changes depending on the format?
That is normal and the prompt expects it. It asks the assistant to report disagreement between samples. You can keep one core profile and add a short format note at the top of each session, or build two profiles if the gap is wide, such as fiction and work email.
Is this the same as the Voice Sheet download?
Same dimensions, different method. The Voice Sheet is a worksheet you fill in by hand. This prompt has the assistant draft it from samples. Many people do both: run the prompt, then correct the result on the sheet.