A full teardown of how we went from 57 posts to a working voice prompt in 4 hours, anonymised with permission pending. You can copy the template at the bottom.
A voice prompt is 500 to 800 words of mechanical rules that make an AI tool write like you. The trick is capturing the rules from your existing posts before writing the prompt, not after. Here's how we did it for a 20k-follower design founder in April, the exact patterns we found, and the template you can copy.
You have probably written a LinkedIn post and thought halfway through: this is taking too long.
The founder we built this voice prompt for had that exact thought in early 2026. He had been using ChatGPT for six months. LinkedIn had quietly stopped giving his AI-drafted posts the reach they used to get. He switched back to writing manually. Every post took 90 to 120 minutes. He runs a design tools company. He does not have 90 to 120 minutes per post.
When we first messaged him he said: Not interested if you are trying to pitch me something.
So we did not pitch. We ran the voice system on his last 57 posts anyway and sent him the result as a free demo.
He replied: I would love to check out the product when it is ready.
This is the teardown of what we sent him. What was in it, what we found in his writing, the sample output we generated, and the voice prompt template you can copy for yourself.
Most people's relationship with ChatGPT looks like this: open a fresh conversation, type Write a LinkedIn post about X, get a generic draft, rewrite half of it, post.
That workflow fails two ways. First, the output sounds like the default voice of the model. Not yours. Second, you end up rewriting anyway, so the time savings are smaller than they should be.
A voice prompt inverts this. You paste 500 to 800 words of mechanical rules into the top of a fresh conversation. Those rules tell the model exactly how you write: average sentence length, phrases you use, phrases you avoid, how you open, how you close, how your tone shifts between different types of content. Then you type the task.
The output comes back in your voice, on first pass, 70 to 90% of the time. The other 10 to 30% you revise. Total time per post drops from 90 to 120 minutes to 15 to 20 minutes.
The prompt is not a template in the sense of fill-in-the-blank marketing copy. It is a config file. Specific. Mechanical. Paste once per chat session and the model behaves differently for the rest of that session.
The input was 57 LinkedIn posts, pulled from the founder's public profile. We did not ask him for anything. The profile was public. The posts were public.
We ran a structured analysis across six dimensions.
Sentence length distribution. We sampled 30 sentences across his posts and counted words per sentence. Average: 7.6 words. That is notably short. Most LinkedIn content sits at 12 to 18 words per sentence. His rhythm was punchier than average, which is a voice signal we needed to preserve in the prompt.
Paragraph structure. Single-sentence paragraphs dominated. 65% of his paragraphs were one sentence. 25% were two sentences. Only 10% were three or more. This is distinctive. It gives his writing a vertical, scannable look on LinkedIn. Any voice prompt trying to match his style had to instruct the model: default to 1-sentence paragraphs.
Opening patterns. We categorised every opening into one of seven archetypes. Forty percent of his posts opened with a personal action ("I pushed some new updates to X this morning"). Twenty percent opened with a contrarian claim ("Most people think X. It is not."). Fifteen percent were announcements, 10% direct advice, 10% personal stories, 5% questions. The breakdown matters because the prompt needs to instruct the model to use the right archetype for the right content type.
Closing patterns. Fifty percent of posts ended with hashtags only. Twenty percent ended with a soft question ("Curious how others are approaching this"). Fifteen percent ended with a declarative mic-drop line ("Accessibility isn't a checklist. It's the baseline."). Fifteen percent ended with a link CTA. This also goes into the prompt.
Signature vocabulary. Phrases he used repeatedly across multiple posts: "design system(s)", "building", "as a designer", "so much more", "I've been", "free". Phrases he never used: "leverage", "synergy", "deep dive", "game-changer", "Stop scrolling", "Agree?". The absence list matters as much as the presence list. A voice prompt should tell the model what to avoid, not just what to include.
Signature moves. We identified five recurring rhetorical patterns:
Each move got cataloged with direct quote examples. The prompt instructed the model to rotate through these moves rather than defaulting to generic structure.
Here is the rough structure of the 720-word prompt we built. Not the actual copy (that is anonymised per the prospect's request), but the sections and what each contains.
Section 1: One-sentence voice summary. Something like: A designer talking to other designers the way he would talk to a friend at a coffee shop, zero pretension, short sentences, heavy line breaks.
Section 2: Mechanical rules. Numbered list. Examples:
Section 3: Signature moves with examples. Named each move and gave 1 to 2 direct-quote examples so the model had something to pattern-match against.
Section 4: Tone by context matrix. Five content types (thought leadership, build-in-public, community, personal, career advice) with 2 to 3 sentences each about how the voice shifts.
Section 5: Vocabulary guidance. Words/phrases to use: 15 signature phrases with example sentences. Words/phrases to avoid: banned words by category.
Section 6: Anti-patterns. Specific things the voice never does. "Never uses Agree? as a closing." "Never uses bold or italic formatting." "Never tags people for engagement."
Section 7: Closing instruction. "Using the rules above, write the following in this voice: [TASK]"
Plain text, no markdown formatting. Copy-paste friendly. Lives as a .txt file the founder owns forever.
We generated five sample posts from the prompt, running on topics we picked from his feed's typical territory. Raw output, no human polish, so he could see what the system actually produces before committing.
The output had: average sentence length of 8 words (within 1 of his actual). Single-sentence paragraphs dominant. His signature phrases showing up naturally. Zero em dashes, zero hedging language, zero "let's dive into". Each post had one of his five signature moves baked in.
Two of the five posts were good enough to publish unedited. Two needed minor tweaks (one sentence each). One needed a regeneration because we picked the wrong content type.
That hit rate is typical for version 1 of a voice prompt. By version 3 (after one round of recalibration) the hit rate climbs to 4 out of 5 usable on first generation.
He went from not interested if you are pitching to I would love to check out the product when it is ready in one reply.
The shift was not to our pitch. We never pitched. The shift was to seeing his own voice reflected back as a working system. The analysis was his voice. The samples were his voice. The prompt was transferable to any AI tool he used going forward.
He had not heard of us before. He had a public stance against AI content. He reversed that stance after one deliverable.
If you want to run this on your own writing, the full method is in The Voice System Playbook. Thirty pages, fourteen chapters, every prompt copy-paste ready.
The shortest version: gather 30 to 50 of your posts, run the voice analysis prompt (Chapter 4 of the Playbook), run the voice prompt assembly prompt (Chapter 5), test the output against 10 samples, iterate once. You will have a working voice prompt in 2 to 3 hours.
If you want us to do it for you instead, The Voice Build is $497 at founder pricing (first 5 buyers) and ships in 3 working days. Everything in this teardown, plus a custom GPT, plus a hook library, plus ten more bonus assets. The comparison: less than 8 months of Pressmaster Pro ($59/mo), and you own every asset forever.
No pressure either way. The Playbook is free and complete. The DFY exists for the afternoon you realise you would rather not spend the weekend.
The Voice Build is $497 at founder pricing (first 5 buyers). Standard price $997. We run the full method on your posts and deliver voice prompt, custom GPT, brand guide, 100+ hooks, 30-day generator, plus 9 more assets in 3 working days. Full refund if the voice is off after revision, and you keep every asset.
See The Voice BuildThirty to fifty is the sweet spot. The minimum that works is about 15 to 20 if the posts are consistent in style. Under 10 posts the analysis starts to describe generic patterns rather than your specific voice signature. If you have fewer than 10 posts, The Brand Build is a better fit because it is interview-led rather than post-analysis-led.
Partially. The voice analysis extracts patterns from text. If your LinkedIn is 80% carousels, you have less text to train on. We supplement by analysing captions, cross-posting text from X or blog, and running shorter analysis on the text you do have. Output quality at that point is 70 to 80% of what it would be with a text-heavy profile.
Custom instructions live permanently in your ChatGPT account and apply to every conversation. A voice prompt is session-specific and much more mechanically detailed. Voice prompts run 500 to 800 words; most custom instructions people write are under 200 words. The specificity is what makes the voice capture work. A custom GPT (covered in The Voice Build) essentially bakes the voice prompt into a dedicated bot so you do not paste it every time.
Yes. You own it. The prompt is a .txt file. Paste it into any tool, hand it to any freelancer, post it in a Notion doc for your team. The only thing we ask is that you do not resell the prompt itself as a product. Everything else is yours forever.