Build Your Brand Voice in 4 Layers: The AI Voice Document System

The problem with "casual but professional"

Most brand voice documents describe adjectives. Casual. Professional. Witty. Approachable. These words mean nothing to an AI model — or to a writer, for that matter. You can't produce consistent output from a list of vibes.

A real voice document describes patterns. What you say repeatedly when you're not thinking about it. What words you never use. How your sentences move. Those are things a model can actually match.

Layer 1: Positions (the hills you'll die on)

Positions are not topics. They're polarising beliefs you hold about your topic — statements that someone could agree or disagree with.

"Taste beats tools" is a position. "AI is changing content creation" is a topic. One gives an AI model something to work with. The other produces generic output.

Find your positions by looking at what you keep saying across your content whether you intended to or not. Three to five of those recurring beliefs are your position set.

Layer 2: Language (always and never lists)

Two lists. Words and phrases you reach for naturally — the ones that feel distinctly yours. And words you clearly avoid — the ones that would sound wrong coming from you.

These lists work as hard rules for any AI model generating content in your voice. The model can't drift into hustle-speak if "hustle" is explicitly banned. It can't sound generic if your specific vocabulary is enforced.

Layer 3: Structure (how your writing moves)

Voice isn't just word choice — it's rhythm. How you open a post. Whether your sentences are short and punchy or long and building. Whether you use line breaks heavily or write in paragraphs. How you close.

These patterns can be described as rules. "Always open with a one-sentence claim, never with context or background" is a structural rule. "Sentences average eight words. Never exceed twelve" is a rhythm rule. Both are things an AI model can follow.

Layer 4: Edge

Edge isn't something you design. It emerges when layers 1 through 3 are right.

If your positions are sharp enough to create disagreement, your word rules are specific enough to be unusual, and your structural patterns are consistent enough to be recognisable — you have an edge. You can't manufacture it by adding it as a fourth layer. It shows up when the other three are working.

Prompt 1 — Find your positions

Start with your last 10–15 posts. This prompt looks for what you keep saying whether you planned to or not.

promptCopy
Here are my last [10/15] posts: [paste them].

Analyse them for recurring opinions, beliefs, and positions. Not topics. Positions. Find the things I keep saying whether I plan to or not.

List them as short, direct statements. Reject anything generic or obvious.

Prompt 2 — Map your language

Run this on the same set of posts immediately after Prompt 1. It builds the always/never word lists.

promptCopy
Now analyse the language in those same posts. Build two lists for me.

ALWAYS: the words and phrases I reach for repeatedly. The ones that feel distinctly mine.
NEVER: the words and phrases I clearly avoid. The ones that would sound wrong coming from me.

Be specific. Single words and short phrases only. No explanations.

Prompt 3 — Lock your structure

The third prompt extracts the patterns in how you write, not just what you write about.

promptCopy
Now analyse the structure of those posts. Find my patterns:

- How do I typically open?
- How do my sentences move?
- How do I typically close?
- What's my average sentence length?
- Do I use line breaks heavily or write in paragraphs?

Write these patterns as rules I could give to an AI so it matches my rhythm.

Prompt 4 — Build the voice document

Run this after the first three prompts have generated their outputs. It consolidates everything into one page you can paste directly into any AI project's custom instructions.

promptCopy
Now write me a one-page voice document I can paste into a Claude project's custom instructions.

Include:
- My positions (from prompt 1)
- My word rules as always/never lists (from prompt 2)
- My structural patterns as rules (from prompt 3)
- 3 behaviour rules:
  (1) never break the word rules
  (2) match my rhythm on every draft
  (3) if something sounds generic, rewrite it until it sounds like a person with a point of view

Keep it tight. One page maximum. No explanations. Just the rules.

How to use the document once you have it

Paste it into the Custom Instructions or system prompt of your Claude project. From that point, every caption, carousel structure, and email draft Claude generates will work from your actual voice patterns — not a generic approximation of what you might sound like.

Revisit it every three to six months. Voice evolves. Run the prompts again on your most recent posts and update the document with anything that's shifted. The document should reflect who you are now, not who you were when you first wrote it.

The 5-point voice checklist

• Positions present — the draft expresses at least one position from the document. It has a point of view, not just information.
• Word bans clear — no words from the NEVER list appear anywhere in the draft.
• Vocabulary present — at least some of the ALWAYS words or phrases appear naturally, not forced.
• Rhythm matches — sentence length and line break patterns match the structural rules.
• Gut test — read it aloud. Does it sound like something you would actually say? If not, rewrite before posting.

Frequently asked questions

What if I don't have 10–15 posts to analyse yet? Use whatever you have — even five posts. The positions and language prompts will return less certain results with a smaller sample, but they'll still surface genuine patterns. Supplement with screenshots of comments you've written, voice notes transcribed, or DMs you've sent. The goal is any authentic output that represents how you actually communicate, not a curated highlight reel.

Do I need a separate voice document for every platform? Not usually. Your core positions, word rules, and rhythm are consistent across platforms — you're the same person whether you're on LinkedIn, Instagram, or writing email. What changes is format and length, not voice. A single voice document applied across platforms will produce content that feels like you everywhere. If you find that a platform genuinely requires a different register, add a short platform-specific section to the document rather than creating a separate one.

How is this different from just giving Claude a few example posts? Examples tell Claude what your content looks like. A voice document tells Claude what rules your content follows. The difference matters at scale: examples lead to imitation that drifts over time, especially as topics change. A document of extracted rules leads to consistent application even on topics you've never written about before. Running the four prompts to build the document is slower than pasting examples, but the output quality is meaningfully better once it's done.