TL;DR
Feed an AI model 8-12 of your best recent posts and ask it to describe (not invent) your voice: opening patterns, sentence-length distribution, phrases you reach for, and the moves you avoid. Turn that analysis into a one-page style doc, then store the doc where it runs on every draft automatically: a Claude Custom Style, a ChatGPT Custom GPT or Project, or a Gemini Gem. The rule that keeps it honest is “archaeology, not aspiration” — every adjective must be backed by a quoted line from a numbered sample, or it gets deleted.
The task
You ship content across LinkedIn, X, a newsletter, maybe a podcast, and the voice drifts. A LinkedIn post reads like a press release; a newsletter goes off in a direction that doesn’t sound like the rest. You want a single style doc you can paste at the top of any draft prompt to pull every channel back to the same voice, and you want that doc built from your actual best posts, not from “bold, authentic” generic adjectives a model made up.
Where AI helps, and where it does not
AI is genuinely good at the archaeological work: pattern-matching across 8-12 samples to find your sentence-length distribution, recurring openings, and phrases you keep reaching for. It is also good at naming opposites (bold vs hedging, warm vs clinical), which is what actually makes the doc useful. Voice without contrast is wallpaper.
AI is bad at imposing a voice you have not actually used. If it suggests “your voice is bold and authentic,” that is a placeholder it generates when your samples were too thin or too uniform. Push back: “name the specific opening pattern from sample 3 and the closing pattern from sample 7.” Distilled voice should be archaeology, not aspiration. If it doesn’t appear in your samples, it isn’t your voice yet.
What to feed the AI
- 8-12 pieces of your best on-brand content from the last 6 months (not 5; thin samples produce generic adjectives)
- The platform/format of each piece (LinkedIn vs newsletter vs thread changes natural cadence)
- The audience you serve, with 1-2 sentence segmentation (e.g., “founders 0-1, technical, dislike hype”)
- 2-3 pieces from creators you do NOT want to sound like — voice is defined as much by what you avoid as what you pick
- 1-2 posts that failed (that you wrote, regretted, and want to flag as off-voice)
- The voice you suspect is yours (it’s a hypothesis the AI will confirm or correct)
- Any banned phrases (corporate-speak, hype words) you already know you don’t use
A note on context limits as of June 2026: ten medium posts plus a few counter-examples is roughly 6,000-12,000 words, which fits comfortably in every current model. Claude Opus 4.7 and Sonnet 4.6, Gemini 3.1 Pro, and GPT-5.5 all carry a 1M-token window (ChatGPT Plus exposes about 320 pages in-app; the full 1M is only on the $200 Pro tier). So you don’t need to summarize your samples — paste them whole, because the model spotting an exact repeated phrase is the entire value.
Copy-ready prompt
Analyze these 10 pieces of my content for voice. Treat this as archaeology, not aspiration — only describe patterns that actually appear in the samples.
Output:
- 3 voice adjectives + 3 explicit opposites (the things I'm not)
- 5 phrases I use a lot — quote exact wording, mark which samples
- 5 phrases that crept in but don't fit my dominant pattern
- Sentence-length distribution (mean, median, longest, % single-clause)
- Formality level (1-10) with one quoted sentence as evidence
- How I open posts (3 dominant patterns) vs how I close (3 dominant patterns)
- 3 structural moves I make that most creators in my niche don't
- A 50-word "if you imitate me" briefing a ghostwriter could pass
Samples:
[paste 10 pieces, separated by ---]Shorter variant — pressure-test an existing style doc
Below is my current voice doc, plus 3 recent drafts. Score each draft 1-10 on voice fidelity. For any score under 8, quote the exact sentence that breaks voice and rewrite it in-voice. Then suggest 1 addition to the doc that would have caught this drift.
Voice doc: [paste]
Drafts: [paste 3, separated by ---]Sample output
A useful pattern-named line: “Opening pattern (5 of 10 samples): start with a concrete artifact, not a claim — e.g., ‘I just shipped X and immediately regretted Y’ instead of ‘Shipping is hard.’ Your hedge-then-commit rhythm is your fingerprint.”
A useful “what you’re not” line: “You are NOT a hype-cycle voice. Banned moves your samples never make: superlatives without numbers, ‘game-changer’, enumerated value-stacks, emoji walls, exclamation marks more than once per post.”
Where to store the finished doc
A style doc only works if it runs on every draft. Pasting it manually each time is the step everyone skips by week two. As of June 2026, each major assistant has a place to store a reusable voice so it applies automatically:
| Tool | Where the voice lives | How to load your doc | Notes |
|---|---|---|---|
| Claude (Opus 4.7 / Sonnet 4.6) | Custom Style, or a Project | Create & edit styles → build from a writing sample, or paste the doc into Project instructions | Project instructions + account Instructions + Style stack on every reply. Styles are migrating to Skills in 2026; the Concise/Explanatory/Formal presets are being removed, so build a custom style rather than relying on defaults |
| ChatGPT (GPT-5.5) | Custom GPT or Project | Paste the doc into the GPT’s Instructions; for a long brand bible, attach it as a Knowledge file so it’s retrieved, not stuffed into context | The “Base style and tone” dropdown (Professional, Candid, Efficient, etc.) is a blunt starting point; your doc overrides it |
| Gemini (3.1 Pro) | Gem, plus Saved Info | Create a Gem and paste the doc into Instructions; add persistent tone preferences under Saved Info | Saved Info applies across all chats; a Gem is the saved-assistant version for one brand |
Pick one home, store the doc there, and stop pasting. If you write in more than one tool, keep the canonical doc in a plain text file and copy it into each — don’t let three drifting versions exist.
How to refine
- Force quoted evidence: “Every adjective must be backed by a quoted line from a numbered sample. Unsupported adjectives are deleted.”
- Pin the opposites: “For each voice adjective, give the opposite I sound like when I drift, not the dictionary antonym. ‘Warm vs corporate’ is more useful than ‘warm vs cold.’”
- Surface the failure mode: “Add a section: ‘When does my voice break?’ Use the failed samples to name 2-3 conditions (rushed, ghostwritten, B2B-ish topic).”
- Make it executable: “Rewrite the doc as a checklist a draft must pass: open with X, sentence length under Y, never use Z. 8 items max.”
- Generate the test post: “Now write a 200-word post in this voice on a topic you’ve never seen me cover. I’ll judge whether it passes the ‘did I write this?’ test.”
Common mistakes
- Defining voice from too few samples. 5 produces generic adjectives; 10 produces patterns
- No contrast section. Voice without “what I’m not” doesn’t survive a busy week
- Letting AI suggest aspirational voice. If it doesn’t appear in your archive, you haven’t earned it
- Updating the style doc but never storing it where it runs automatically (a Style, GPT, or Gem)
- Treating voice as fixed. It shifts with audience changes, after a major project, post-burnout
- Optimizing voice to be identical across channels instead of letting one fingerprint translate into different formats
- Skipping the failure-mode section. Knowing when your voice breaks is more useful than knowing what it is
- Building voice from “wow this resonated” posts only. The high-engagement piece may be off-voice but algorithmically rewarded
FAQ
- How many samples do I need?: 8-12 of your best, recent posts. Under 5 is too thin; over 15 starts averaging your evolution into mush. All of that fits in the 1M-token window every current model carries, so paste them whole.
- Should I update the style doc?: Every 6 months, or after a major audience shift, niche pivot, or 3 months post-burnout (your voice changes after long breaks). Re-run the analysis on the last 90 days only.
- Can AI write in this voice automatically?: Yes. Store the doc as a Claude Style, a ChatGPT Custom GPT, or a Gemini Gem so it loads on every draft. Quality usually lands around iteration 3; the first AI-voiced post is often 60% there.
- Which model is best for this?: Any of the top three handles the analysis well. Claude Opus 4.7 and Sonnet 4.6 tend to be strong at long-form voice mimicry; GPT-5.5 and Gemini 3.1 Pro are close. The bigger lever is sample quality, not model choice.
- What if my voice doesn’t match the niche?: Good. Voice is a moat. The “this person sounds different” reaction is the entire point. Only worry if readers say “this sounds wrong,” not “this sounds different.”
- Should I have different docs per platform?: One core voice doc plus a 50-word per-platform delta (LinkedIn: longer hooks; X: harder line breaks; newsletter: longer paragraphs). Don’t fragment into 4 voice docs.
Related
- Personal brand prompts — broader personal brand prompts
- Brand story prompts — story-shaped brand
- Brand voice definition prompts — formal brand voice doc
- Brand tone guide AI — turn voice into tone guide
- Personal brand statement — broader positioning
- Content calendar creator AI — produce in this voice
- Profile bio — bios in this voice
Tags: #AI writing #Content creation #Creator #Personal brand #Brand voice