Ghostwrite Yourself

You can feel it in the second paragraph. The hedging, the “moreover,” the symmetric three-bullet lists, the closing sentence that says nothing. You spend more time deleting AI tics than you would have spent writing the thing yourself.

The fix isn't a better prompt. It's a better profile: a single document that captures how you actually write, what you refuse to write, and who you write for. You build it once. You paste it into every system prompt, every Project, every custom GPT, every Claude session, from then on.

Most “voice clone” tutorials get this backwards. They hand you a hundred questions and ask you to describe your own voice from memory. People are bad at that. They write down the voice they wish they had, then wonder why the output sounds aspirational and weird.

The result is a 3,000–4,000 token markdown file that you'll keep for years. Update it when your taste shifts. Port it to whatever model comes next. It's the most leveraged hour of work you'll do this quarter.

Time investment

~1 hourto gather samples, answer 22 questions, compile the file

Output

One .md fileportable across Claude, ChatGPT, Gemini, Grok, anything

Decay rate

~6 monthsrefresh when your voice or audience meaningfully shifts

A great voice file has three ingredients in a strict priority order. Most guides ignore the first two and over-index on the third. That's why the output sounds generic. Get the order right and the rest is mechanical.

Answer in your own words. Be specific. If a question doesn't apply, write why. That's signal too. Print and fill, or copy into a doc.

Use this if you'd rather dictate your answers than fill the worksheet. Paste it into Claude (Opus, with extended thinking on if you've got it). Speak your answers using Wispr Flow or your phone's dictation. The Interviewer will push back on vague answers and dig deeper.

You are interviewing me to build a profile of how I write and think. I'll be dictating, so my
answers may ramble. That's fine, capture everything.

Ask me the 22 questions below one at a time. After each answer:
  - If I was vague, gave a one-word reply, or contradicted myself, push back and ask a sharper
follow-up before moving on.
  - Don't accept "I don't know" without trying once to unstick me.
  - Don't summarize as we go.

When all 22 are answered, just say "done" and wait. Don't analyze yet.

THE 22 QUESTIONS:
Audience & Purpose
1. Who are you writing for? Describe one specific real person if you can.
2. What do you want them to feel, think, or do after reading?
3. What's the one thing you want to be known for saying?
Voice & Tone
4. If a friend read your writing without your name on it, how would they know it was you?
5. How casual or formal are you? Where's the line you don't cross?
6. Do you swear? When, and when not?
7. Funny, serious, warm, sharp, dry: pick the two or three that actually fit you.
8. What words or phrases make you cringe when other writers use them?
Refusals & Pet Peeves
9. What topics or angles will you never write about?
10. What kind of advice or content do you refuse to give, even when asked?
11. What does generic AI writing get wrong that drives you up the wall?
Style & Structure
12. Long sentences or short? Paragraphs or one-liners? Lists or prose?
13. How do you open a piece? How do you end one?
14. Do you use headers, bullets, bold? When do they help vs. clutter?
15. How much do you explain vs. assume the reader keeps up?
Influences & Anti-Influences
16. Name 2-3 writers/creators whose voice you admire. What specifically?
17. Name 1-2 whose style you actively reject. Why?
18. If your writing had a soundtrack genre, what would it be?
Subject & Stance
19. What are the 3-5 topics you actually write about?
20. What's a take you hold that most people in your field disagree with?
21. What do you wish more people understood about your subject?
22. What's the difference between writing you're proud of and writing that just got the job done?

Begin with question 1.

This is the prompt that does the real work. Paste it into a fresh Claude session, then paste your writing samples, transcripts, and 22 answers below it. The “trust the samples over self-report” instruction is the line that prevents the profile from becoming aspirational fiction.

Below is everything you need to build my voice profile: writing samples, transcripts (optional),
and my answers to a 22-question interview.

Compile this into a single markdown file (~3,000-4,000 tokens) that another instance of you could
load to write in my voice.

Structure it as:

1. VOICE FINGERPRINT
   5-8 bullets that capture what makes my writing mine.
   Derived from the samples first, the answers second.

2. AUDIENCE & PURPOSE
   Who I write for, what I'm trying to do.

3. DO
   Concrete moves, phrasings, structural habits, tonal range.

4. DON'T
   Refusals, banned phrases, tics to avoid, AI tells I hate.

5. REFERENCE EXAMPLES
   3-5 short excerpts from my actual samples that demonstrate the voice,
   each with a one-line note on why it's characteristic.

6. CALIBRATION NOTES
   When to dial casual up/down. When to swear. When to be serious.
   When to be playful. Edge cases.

CRITICAL RULES:
  - Prioritize the writing samples over my self-description.
  - If what I SAY I do contradicts what the samples SHOW I do, trust the
    samples and flag the gap explicitly.
  - Be specific. Avoid generic style descriptors like "engaging,"
    "authentic," "conversational". Those are useless.
  - Quote actual phrases I use. Name actual structural moves I make.
  - If a category is thin (e.g. only 2 samples), say so. Don't fabricate.

INPUT MATERIALS:

[paste writing samples here]

[paste transcripts here, if any]

[paste 22 answers here]

The hour you spend on this is worth more than any prompting course. Most people skip it because the payoff is invisible until the first time you paste your profile into a blank session and the output already sounds like you. After that, you don't go back.

The profile isn't magic. It's a portable filter. It won't think for you. It won't have your judgment. It will not catch the things only you would catch. What it will do is collapse the gap between the AI draft and the version you'd actually publish, so you spend your time on the parts that need a human, not deleting “moreover” and “delve” out of the rest.

Build it once. Deploy it everywhere. Update it when your taste shifts. That's the whole game.