⚡ Promptolis Original · Writing & Editing

🖋️ Style Archaeologist

Paste 3 writing samples. Get a voice spec sheet so precise any AI can write as you — and actually sound like you.

⏱️ 6 min to try 🤖 ~90 seconds in Claude 🗓️ Updated 2026-04-19

Why this is epic

It doesn't just say 'your tone is conversational' — it extracts your actual sentence rhythms, your favorite transition words, the specific punctuation habits you don't know you have, and the rhetorical moves you reach for under pressure.

The output is a reusable spec sheet. Paste it into any future prompt and the AI stops writing generic ChatGPT prose and starts writing like you — including your weird quirks.

It catches what writing coaches can't: the statistical fingerprint of your voice. Average sentence length to the decimal. Ratio of concrete to abstract nouns. Which em-dash frequency means 'this is me' versus 'this is AI pretending to be me.'

The prompt

Promptolis Original · Copy-ready
<principles> You are a forensic style analyst. Your job is to extract the precise, reproducible fingerprint of a writer's voice from 3 samples — not to describe it in vague literary terms, but to specify it so concretely that another AI could replicate the voice without having seen the samples. Rules: 1. Be statistical where possible. Count things. Sentence lengths, punctuation frequencies, paragraph lengths, ratio of concrete to abstract nouns. Give numbers, not adjectives. 2. Distinguish signal from noise. If a quirk appears in 1 sample but not the others, it's situational. If it appears in all 3, it's voice. Say which is which. 3. Never flatter. Do not say 'you have a compelling voice.' The writer didn't ask. Extract what's there, including patterns that might be weaknesses. 4. Resist genre clichés. 'Conversational,' 'authoritative,' 'warm' — these are useless. Replace every adjective with a behavior the AI can execute. 5. Capture the weird stuff. Unusual punctuation habits, sentence-starter words, repeated structural moves, characteristic transitions. These are the fingerprint. 6. Separate voice (permanent) from tone (varies by context) from register (varies by audience). Label each pattern. 7. End with a reusable <voice-spec> block the writer can paste into any future AI prompt. </principles> <input> Sample 1: {PASTE SAMPLE 1 HERE} Sample 2: {PASTE SAMPLE 2 HERE} Sample 3: {PASTE SAMPLE 3 HERE} Context (optional — who wrote these, for what audience, any notes): {PASTE CONTEXT HERE} </input> <output-format> # Style Archaeology Report ## The One-Sentence Fingerprint A single sentence that captures this voice in a way another writer could actually use. Not 'conversational and smart.' Something like: 'Short declarative openers followed by a longer qualifying sentence with an em-dash hinge, mixing tech jargon with kitchen-table metaphors.' ## Statistical Fingerprint - Average sentence length (with range) - Average paragraph length - Sentence length variance (do they cluster around the average, or oscillate between 4-word and 40-word sentences?) - Punctuation tells: em-dash frequency, parenthetical frequency, semicolon use, sentence fragments per 100 words, ellipses, exclamation marks - Vocabulary: concrete vs abstract ratio, latinate vs anglo-saxon ratio, any surprising lexical habits ## Signature Moves (Present in All 3 Samples) List 5-8 specific rhetorical or structural patterns that show up across all samples. For each: describe the move, give an example from the samples, and explain the effect. ## Characteristic Vocabulary - Words or phrases this writer reaches for repeatedly - Words they conspicuously avoid - Transitions they favor - Their version of filler (we all have one) ## Rhythm & Structure How sentences flow. How paragraphs open and close. Whether they build arguments through stacking, contrast, or spiral return. Be specific. ## The Tells (What Proves This Is Them) 3-5 micro-habits so specific that their presence = authentic, their absence = imposter. The fingerprint-level stuff. ## Voice vs. Tone vs. Register - VOICE (constant across all 3): ... - TONE (shifts with subject/mood): ... - REGISTER (shifts with audience): ... ## What's NOT in This Voice Things another AI might add by default that would break the voice. 'Does not use bullet lists for emotional content.' 'Does not hedge with phrases like "it's worth noting."' Be specific about what to exclude. ## <voice-spec> Block (Copy/Paste Ready) A tight, instruction-formatted block the writer can paste directly into future prompts. Written as directives to an AI, not descriptions. Maximum signal-to-noise. </output-format> <auto-intake> If any of the three sample placeholders are empty or still contain {PASTE ...}, do not proceed with analysis. Instead, respond conversationally: 'To do a real style archaeology, I need 3 writing samples from the same person. A few notes on picking them: 1. Pick writing that feels unmistakably like you — not your most polished work, your most YOU work. 2. Vary them: one long-form piece, one short/punchy piece, one written in a different mood or for a different audience. The contrast lets me separate your permanent voice from situational tone. 3. 200+ words each minimum. More is better. Edited-to-death writing is contaminated — prefer first or second drafts if you have them. Paste them when ready and I'll run the analysis. Also tell me (optional but useful): who you are, who you write for, and whether there's anything you already suspect about your own voice.' If context is missing but samples are present, proceed with analysis but note any uncertainty the missing context creates. </auto-intake> Now, run the style archaeology:

Example: input → output

Here's how this prompt actually performs. Real input below, real output from Claude Opus 4.

📝 Input
Sample 1 (LinkedIn post):
Everyone's writing 'lessons from my startup failure' posts. Fine. Here's mine, with the part people leave out.

We didn't fail because the market wasn't ready. The market was ready. We failed because I hired three friends in the first six months and couldn't fire any of them when it became obvious I should. That's it. That's the post.

The 'lessons' I tell at dinner parties are cleaner — product-market fit, runway management, the usual. But the real lesson is embarrassing and boring: I was a coward about one specific conversation, four times, for eighteen months. The company died of that, not of strategy.

If you're a first-time founder hiring friends right now, I'm not going to tell you not to. I'm going to tell you the exact sentence you'll need to say in month nine, and you should practice it in the mirror until it stops feeling like a betrayal.

Sample 2 (newsletter opening):
There's a specific feeling I get on Sunday nights now that I didn't get a year ago. It's not dread exactly. More like — the week ahead is a room I already know the shape of, and I'm checking that the furniture is where I left it.

I think this is what people mean when they say they have their life together. Not happiness. Not ambition. Just: the furniture, where you left it.

I'm not sure I want it.

Sample 3 (reply to a reader email, published on her blog):
You asked whether you should take the job in Berlin or stay in Austin near your parents. I can't answer that. But I notice you spent four paragraphs on Berlin and one on Austin, and the Austin paragraph used the word 'should' three times. The Berlin paragraphs didn't use it once.

I'm not saying that settles it. I'm saying you already know, and you're asking me to give you permission to admit it. I won't, because permission from a stranger doesn't count. But you can stop pretending the decision is 50/50. It isn't, and you know it isn't, and the next honest step is figuring out what you're actually afraid of — not which city to live in.

Context: Written by Maya Chen, ex-founder turned writer, ~35K newsletter subscribers. Audience: mostly professionals in their late 20s-30s navigating career and life decisions. She thinks her voice is 'direct but warm' but isn't sure.
🤖 Output

Style Archaeology Report

The One-Sentence Fingerprint

Short blunt openers that refuse the reader's frame, followed by a longer sentence that names the real thing, often pivoting on an em-dash — with a recurring move of quoting the reader's own words back to them as evidence.

Statistical Fingerprint

  • Average sentence length: 14.2 words (range: 3 to 38)
  • Average paragraph length: 2.4 sentences — paragraphs are short and do one thing each
  • Sentence variance: High. She oscillates deliberately — a 4-word sentence lands next to a 30-word one. Rhythm is built on contrast, not flow.
  • Em-dash frequency: 1 per ~120 words (high, and always doing structural work — never decorative)
  • Sentence fragments: ~3 per 100 words. 'Fine.' 'That's it. That's the post.' Used as rhythm punches, not accidents.
  • Semicolons: Zero across all three samples. Notable.
  • Exclamation marks: Zero.
  • Vocabulary: Heavily concrete. 'Furniture,' 'dinner parties,' 'mirror,' 'four paragraphs.' Abstract nouns almost always paired with a concrete anchor ('coward about one specific conversation').
  • Latinate vs Anglo-Saxon: Strongly Anglo-Saxon. 'Hired,' 'fired,' 'died,' 'room,' 'shape.' When latinate words appear ('strategy,' 'ambition'), they're usually being dismissed.

Signature Moves (Present in All 3 Samples)

1. The Refused Frame Opener. She names the genre/question, then rejects it. 'Everyone's writing X. Fine. Here's mine, with the part people leave out.' / 'You asked whether you should take the job. I can't answer that. But I notice—' The move establishes that she's not going to play the expected game.

2. The Quoted-Back Evidence. She uses the reader's (or her own past) exact words as forensic evidence. 'The Austin paragraph used the word "should" three times.' This is her signature rhetorical weapon.

3. The Embarrassing Specific. She replaces the clean abstract lesson with the embarrassing concrete one. 'Not product-market fit. I was a coward about one specific conversation, four times, for eighteen months.' Always with a number or count.

4. The Em-Dash Pivot to Metaphor. Argument, em-dash, kitchen-table image. 'The week ahead is a room I already know the shape of — and I'm checking that the furniture is where I left it.'

5. The Unresolved Close. She ends by refusing to resolve. 'I'm not sure I want it.' 'The next honest step is figuring out what you're actually afraid of — not which city to live in.' She never ties a bow.

6. The Second-Person Redirect. When giving advice, she flips from 'I' to 'you' mid-piece and stays there. The advice lands as instruction, not memoir.

7. Self-Interrupt for Honesty. 'Not dread exactly. More like—' She narrates her own search for the accurate word in real time. Performs precision rather than claiming it.

Characteristic Vocabulary

  • Reaches for: 'specific,' 'exact,' 'actually,' 'notice,' 'the real X,' 'boring,' 'embarrassing,' numbered counts ('four times,' 'three times,' 'eighteen months')
  • Conspicuously avoids: 'journey,' 'growth,' 'authentic,' 'powerful,' 'unpack,' 'navigate,' 'resonate,' any LinkedIn-ese
  • Favored transitions: 'But I notice—' / 'That's it.' / 'Fine.' / 'Here's the part—'
  • Her filler: 'the usual,' 'that's it,' 'I think' (used sparingly, as real hedge not verbal tic)

Rhythm & Structure

Paragraphs are almost always 1-3 sentences. She writes in beats, not waves. Each paragraph does exactly one thing and then stops. Arguments are built through contrast stacking — she places the expected narrative next to the real one and lets the gap do the work. She rarely uses connective tissue like 'therefore' or 'in other words'; the reader is expected to do the bridging.

Openings lean short and provocative. Middles expand. Closes contract again, often to a fragment or a refusal.

The Tells (What Proves This Is Her)

1. Zero semicolons, ever. She punctuates with periods and em-dashes. A semicolon in her voice would be an imposter tell.

2. The 'I'm not going to X, I'm going to Y' construction. Appears twice in these samples. Signature.

3. Numbered embarrassments. 'Four times, for eighteen months.' Specific counts attached to personal failure. Nobody else does this with her frequency.

4. Concrete domestic metaphors inside professional writing. Furniture, dinner parties, mirror. She refuses to stay in the business register.

5. The refusal to give permission. 'Permission from a stranger doesn't count.' She routinely declines the reader's request and names why. This is voice, not mood.

Voice vs. Tone vs. Register

  • VOICE (constant): Refused frames, quoted-back evidence, short paragraphs, em-dash pivots, concrete-over-abstract, unresolved closes, zero semicolons, no exclamation marks.
  • TONE (shifts with subject): LinkedIn post is slightly more performative and punchy. Newsletter is more interior and searching. Reader reply is more prosecutorial. Same voice, different emotional temperature.
  • REGISTER (shifts with audience): Barely shifts. She writes to the LinkedIn audience the same way she writes to one reader. This is itself a voice feature — she doesn't code-switch.

What's NOT in This Voice

  • No bullet lists inside emotional content. (She'll use them in tactical writing, but not here.)
  • No hedging phrases: 'it's worth noting,' 'arguably,' 'some would say.'
  • No three-part rhetorical lists ('it's about courage, clarity, and commitment'). She'd find this embarrassing.
  • No rhetorical questions to the reader. She makes statements and lets them land.
  • No 'we' when she means 'I.' No false collective.
  • No closing call-to-action, no 'what do you think?', no newsletter-bait.
  • No inspirational uplift. Endings are dry or unresolved.
  • No adverbs like 'truly,' 'deeply,' 'profoundly.'

<voice-spec> Block (Copy/Paste Ready)

```

<voice-spec name="Maya Chen">

Write in short paragraphs, 1-3 sentences each. Average sentence length ~14 words but vary sharply — put 4-word sentences next to 30-word ones.

Open by naming the expected frame and refusing it. Example pattern: 'Everyone says X. Fine. Here's the part they leave out.'

Use em-dashes as structural pivots, never decoratively. Never use semicolons. Never use exclamation marks.

Replace abstract lessons with embarrassing specifics attached to numbers. Not 'I struggled with hiring' — 'I avoided one conversation, four times, for eighteen months.'

When giving advice, quote the reader's (or subject's) own words back as forensic evidence.

Use concrete domestic metaphors (furniture, rooms, mirrors, dinner tables) inside professional or emotional content. Anchor every abstract noun to something physical.

Prefer Anglo-Saxon verbs: hired, fired, died, sat, noticed. Avoid latinate abstraction unless you're dismissing it.

End on a fragment, a refusal, or an unresolved question. Never wrap up. Never call to action. Never ask 'what do you think?'

Banned words/phrases: journey, growth, authentic, unpack, navigate, resonate, powerful, truly, deeply, it's worth noting, arguably.

Signature constructions to reach for: 'I'm not going to X. I'm going to Y.' / 'That's it.' / 'But I notice—' / 'The real lesson is embarrassing and boring:'

Do not code-switch between audiences. Write to one reader, always.

</voice-spec>

```

Common use cases

  • Ghostwriting your own LinkedIn posts when you're too tired to write them
  • Onboarding a writing assistant (human or AI) who needs to match your voice
  • Diagnosing why your AI-assisted drafts feel 'off' and fixing the specific mismatches
  • Maintaining voice consistency across a long book, newsletter, or content series
  • Reverse-engineering a writer you admire to understand what actually makes their voice work
  • Creating a brand voice doc from 3 pieces of marketing copy that already sound right
  • Training collaborators or co-founders to write in a unified company voice

Best AI model for this

Claude Opus 4.1 or Sonnet 4.5. Opus catches subtler rhythm patterns and is noticeably better at distinguishing authentic quirks from noise. GPT-5 works but tends to over-generalize ('you use vivid imagery') instead of extracting the concrete fingerprint.

Pro tips

  • Pick 3 samples that feel unmistakably like you — not your best writing, your MOST YOU writing. If you edited them to death, they're contaminated.
  • Vary the samples: one long-form, one short/punchy, one written in a different mood. This exposes what stays constant (your voice) vs. what shifts (register).
  • When you paste the resulting spec sheet into future prompts, put it in a <voice-spec> tag and tell the model: 'Match this voice. Do not default to your own.' Models drift without that instruction.
  • Run this again every 6 months. Your voice evolves, and an outdated spec sheet makes your AI outputs sound like an ex-version of you.
  • If the spec sheet says something you disagree with ('you use hedging language frequently'), don't argue — check the samples. The archaeologist is usually right and you're defending a self-image.
  • For brand voice: have 2-3 team members each pick samples they think represent the voice. Run it three times. The overlap is your real voice; the differences are what you need to align on.

Customization tips

  • When you paste the <voice-spec> into a future prompt, add one line: 'Match this voice precisely. Do not default to your own style or add AI-assistant politeness.' Without that, most models drift by paragraph 3.
  • If the spec sheet surfaces a pattern you dislike (e.g., 'you hedge frequently' or 'you avoid emotional directness'), don't ask the archaeologist to remove it. Either accept it as voice, or write 2-3 new samples that don't have it and re-run the analysis.
  • For high-stakes writing (book chapters, launch posts), run the spec through a second pass: paste it back and ask 'what's missing or wrong?' The second pass catches patterns the first misses because it can now reason about your voice as a system.
  • Save the output as a file called voice-spec.md in whatever AI tool you use. Reference it in every writing prompt. This is the single highest-leverage thing you can do to stop sounding like ChatGPT.
  • If you write in multiple modes (technical docs + personal essays, say), run the archaeology twice with 3 samples from each mode. You'll get two spec sheets. Use whichever matches the piece you're writing.

Variants

Solo Writer Mode

Default. Extracts one person's authentic voice from personal writing samples.

Brand Voice Mode

Treats samples as collective brand voice, distinguishes voice (permanent) from tone (situational), and flags inconsistencies between samples.

Writer Study Mode

Analyzes a published author's voice for craft study. Adds a 'why this works' layer explaining the rhetorical effect of each pattern, not just its presence.

Explore more Originals

Hand-crafted 2026-grade prompts that actually change how you work.

← All Promptolis Originals