How-To
May 202611 min read

How to Reverse Engineer Your Own Writing Voice (7-Step Discovery Method)

Most people don't need to invent a writing voice. They already have one and don't know what it is. Here is the 7-step process for extracting the voice patterns hiding in your existing writing, then turning them into something an AI can actually use.

You already have a voice. The work is finding it, not inventing it. Pull 10-20 of your existing writing samples, run a 7-step pattern scan, and translate the findings into the five-section voice prompt structure. Time required: 60-90 minutes for discovery, plus an hour to package. Output: a voice prompt that sounds like you because it was built from how you already write.

Why discovery beats invention

Most articles about "finding your voice" treat it as a creative exercise. They ask you to brainstorm adjectives. Punchy. Conversational. Authoritative. The output is a list of words that describes a voice you do not have, and which leaves you no closer to producing content that sounds like you.

Discovery works the other way round. You take writing you have already produced and look for the patterns that make it recognisable. The voice is in there. The exercise is finding it, not making it up. How to build a voice prompt covers the construction phase. This article covers the discovery phase that should usually come first.

The seven steps

Each step builds on the previous one. Don't skip ahead. The pattern hunting is the whole exercise.

Step 1

Gather 10-20 writing samples

Open a blank document. Paste in 10-20 pieces of your own writing. Mix the formats deliberately:

  • 5 LinkedIn posts (or whatever your main public writing is)
  • 3 emails you sent to clients or prospects
  • 2 emails you sent to friends or colleagues (less performative)
  • 2 voice notes transcribed (way you actually talk)
  • 2 pieces of sales or marketing copy if you have it
  • 2-4 wildcards: a Slack message, a long DM, a comment on someone's post, a journal entry

The mix matters. Public writing reveals performative voice. Private writing reveals natural voice. The gap between the two is information.

Done when:
  • You have at least 3,000 words of your own writing in one place
  • The samples cover at least four different contexts
  • None of the samples are co-written or heavily edited by someone else

Step 2

Identify your three best samples

Read everything once. Mark the three pieces that feel most you. Not the ones that performed best. Not the most polished. The ones that, if a close friend read them blind, would be guessed as yours.

These three anchor everything that follows. They are the centre of gravity. Patterns that show up in all three are core voice. Patterns that show up in only one are likely context-specific.

Done when:
  • You have three pieces marked as "most me"
  • You can articulate, in one sentence, why each piece feels like you

Step 3

Run the mechanical scan

Now go technical. Across the three best samples, measure:

  • Sentence length range. Shortest sentence and longest sentence in words. Average. Standard deviation if you want to be exact.
  • Paragraph length. One-line paragraphs? Two-to-four-line? Walls? What is your default?
  • Contractions. Do you use "I'll", "won't", "we're"? Always? Never? Sometimes?
  • Punctuation habits. Em dashes? Semicolons? Ellipses? Lists with bullets or with commas?
  • Common openers. Do you start sentences with "And", "But", "So" often? With participles? With questions?
  • List vs prose. When you have three items, do you write them as a sentence, a list, or paragraphs? What is the trigger?
  • Active vs passive voice. Quick scan. Most natural-voice writers are 90% active.

Write down the numbers. Vague is useless. "Sentences are short" doesn't help an AI; "Sentences range from 4 to 22 words, average 11" is usable.

Done when:
  • You have a numbered profile of mechanical patterns across the three samples
  • You could explain the patterns to someone who has never read your writing

Step 4

Extract banned words

Two passes:

Pass 1: words you systematically avoid. Read all 10-20 samples. Note words that never appear. "Leverage" is a classic absence in good voice. So is "synergy", "ecosystem", "robust", "delve", "tapestry". The ones missing from your writing are the ones that should be banned in the voice prompt.

Pass 2: words you would never write but ChatGPT would. Add the AI-default vocabulary you do not use. The list is well-documented: "leverage", "cutting-edge", "thought leader", "best-in-class", "unlock", "utilise", "transformative", "elevate", "navigate", "streamline", "seamlessly", "in this fast-paced world".

Combined output: 15-30 banned words. Some are personal idiosyncrasies. Some are AI-default. Both belong in the voice prompt.

Done when:
  • You have a list of 15-30 words that should never appear in content claiming to be in your voice
  • The list mixes personal idiosyncrasies with AI-default vocabulary

Step 5

Find your signature moves

This is the part most discovery exercises skip and the part that matters most. Signature moves are the 3-5 distinctive habits that make your writing recognisable. Look for:

  • Recurring framings. Do you tend to set up an idea by stating its opposite first? Do you use "most people think X. The reality is Y" structures?
  • Structural tics. Do you open with a question? Close with a one-line provocation? Use mid-paragraph asides in parentheses?
  • Callback patterns. Do you reference an earlier point in the closing line? Do you book-end posts with the same image or phrase?
  • Recurring metaphors. Do you use sports metaphors, kitchen metaphors, building/architecture metaphors? Most writers have a metaphor lane they default to.
  • Compression habits. Do you write "Three things." then list them? Do you compress complex ideas into a single coined phrase?

Pick the 3-5 strongest. These are what makes your writing yours, more than sentence length or word choice.

Done when:
  • You have 3-5 signature moves named and described
  • Each one has at least two examples from your own writing

Step 6

Map tone shifts by context

Voice is not one thing. It shifts by context. Compare how you sound in:

  • A LinkedIn post (medium-public, professional)
  • A client email (one-to-one, professional, accountable)
  • A friend message (private, casual, no performance)
  • A sales page (asymmetric, promotional, structured)
  • A blog or long-form piece (sustained, layered)

For each context, note: how does sentence length shift? Does humour increase or decrease? Are you more direct or more hedged? Do contractions go up or down?

The shifts are part of the voice. A voice prompt that captures only one register produces content that sounds wrong in other contexts.

Done when:
  • You have a 4-6 row table mapping context to tone shift
  • Each row has a specific change (not just "more formal")

Step 7

Translate findings into the five-section voice prompt

Now package the discovery into the voice prompt structure used across the Syxo system:

  1. Voice essence. One paragraph (60-100 words) describing how the writer communicates. Description, not adjectives. Built from the patterns in Steps 2-3.
  2. Mechanical rules. Sentence length, paragraph length, contractions, punctuation, openers, list vs prose, active/passive. Direct from Step 3.
  3. Banned words. The 15-30 word list from Step 4.
  4. Tone by context. The 4-6 row matrix from Step 6.
  5. Signature moves. The 3-5 distinctive habits from Step 5.

Target length: 500-800 words. Anything shorter loses too much; anything longer overwhelms the AI.

Done when:
  • The voice prompt is 500-800 words
  • Every section is filled with specifics, not abstractions
  • You can read the prompt and see your own voice in it

What the discovery typically reveals

Three patterns show up across most discovery sessions:

Your real voice is sharper than your performative voice. The writing you produce in DMs, voice notes, and casual emails tends to be more direct, more specific, and more opinionated than what you publish. The discovery surfaces the gap. Most people decide to publish in their real voice once they see how much stronger it is.

You have signature moves you have never named. The 3-5 habits in Step 5 are nearly always there before the writer notices them. Naming them turns unconscious habits into reusable patterns the AI can replicate.

Banned words are about half personal and half AI-default. The personal half (the words you avoid) is more interesting than the AI-default half (the words everyone avoids). The personal banned list is what stops a voice prompt sounding like a generic "write professionally" instruction.

What to do with the output

Once the voice prompt is built, the next steps are:

  1. Drop it into ChatGPT. Build a Custom GPT and paste the voice prompt into instructions. How to train ChatGPT on your writing style covers the setup.
  2. Drop it into Claude. Build a Project and paste the same prompt. How to train Claude on your writing covers Claude specifically.
  3. Test it. Use the prompts in best ChatGPT prompts for LinkedIn to generate sample posts and check them against your samples.
  4. Iterate. When the output drifts, find which section of the voice prompt is too vague and tighten it.

For the diagnostic side — testing whether the voice prompt actually works — see how to audit your AI content.

When to do this yourself versus when to outsource

Discovery is doable yourself if:

Outsourcing makes sense if:

The Syxo DFY Voice System runs both phases for you and ships the voice prompt, Custom GPT, Claude Project, and starter content in 2-3 working days.

Related reading

Don't want to do this yourself?

The DFY Voice System runs the discovery and construction phases for you. We analyse 10-20 samples, extract the patterns, and ship a voice prompt, Custom GPT and Claude Project in 2-3 working days. The Voice Build methodology, applied to your existing writing. £497 founder pricing.

See The Voice Build

Frequently Asked Questions

What does reverse engineering your voice mean?

Looking at writing you have already produced and extracting the patterns that make it sound like you, instead of inventing a voice from scratch.

How is this different from building a voice prompt?

Discovery surfaces what your voice actually is. Construction packages the findings into a 500-800 word prompt the AI can use.

How many writing samples do I need?

10-20 samples is the sweet spot. Mix formats: posts, emails, voice notes, sales copy. The mix is what reveals tone shifts.

Can I do this if I haven't written much?

Yes. Supplement with voice notes (transcribed) and recorded conversations. The way you talk reveals voice patterns even when you haven't written them down.

What if I don't like my own voice?

The voice you don't like is often a default mode (corporate, hedged) rather than your real voice. Sample casual contexts (DMs, voice notes). The real voice is usually sharper than expected.

How long does this process take?

60-90 minutes for discovery, plus an hour to translate findings into a voice prompt.