How-To
May 202613 min read

How to Build a Voice Prompt That Actually Works

The five-section structure, real examples, the testing protocol, and the four mistakes that kill voice prompts before they get a chance. Three hours of focused work, output that sounds like you on first draft.

A voice prompt is a 500-800 word reference document with five sections: voice essence, mechanical rules, banned words, tone by context, signature moves. Built right, it produces 70-85% on-voice output on first draft. Built vaguely (which is most of the time), it produces output indistinguishable from default AI. The structure below is the difference.

Why most voice prompts don't work

The most common voice prompt looks like this: "Write in a friendly, professional tone. Be concise. Use a conversational style. Avoid jargon."

This produces nothing. The AI defaults to its baseline voice because the instruction is too vague to operationalise. "Friendly" means something different to every reader. "Concise" has no specification. "Conversational" describes a vibe, not a structure.

A voice prompt that actually works tells the AI specific, mechanical things it can match — sentence length ranges, banned words, signature sentence-opening patterns, tone shifts by format. Specificity is the difference. The five-section structure below is the framework that produces specificity reliably.

The five-section structure

SECTION 1

Voice essence

One paragraph describing how this person communicates. Not adjectives — a description specific enough that someone reading it could write a sentence in the voice without ever seeing samples.

Bad: "Friendly, professional, approachable."
Good: "Talks like a smart friend who's done this before. Direct, specific, slightly opinionated. Uses short sentences for emphasis and longer ones for explanation. Names the reader's pain before naming the solution."

The good version gives the AI directional guidance. The bad version gives it nothing actionable.

SECTION 2

Mechanical rules

The quantifiable patterns the AI can mechanically reproduce. Use ranges (12-18 words) rather than absolutes (15 words) — real human voices vary deliberately for emphasis.

The eight rules to specify:

  • Sentence length range (e.g., "8-22 words, with frequent short sentences for emphasis")
  • Paragraph length range (e.g., "1-3 sentences per paragraph")
  • Contractions (always / never / sometimes — and which contexts)
  • Punctuation preferences (em dashes yes/no, semicolons yes/no, ellipses yes/no)
  • Sentence opening patterns (can start with "And" or "But"? Use rhetorical questions?)
  • List vs. prose preference (when to use bulleted lists vs. inline prose)
  • Active vs. passive voice ratio
  • Number of words for typical post (LinkedIn: 150-300; blog intro: under 150; etc.)

Specifying all eight produces consistent output. Specifying two or three produces inconsistent output that drifts back to default.

SECTION 3

Banned words and phrases

15-30 specific phrases the writer would never use. The list closes the gap between "your voice" and "default AI vocabulary."

Common banned-word categories:

  • AI defaults: "leverage," "cutting-edge," "thought leader," "best-in-class," "synergy," "delve into," "navigate," "streamline," "robust," "seamlessly," "tapestry," "elevate," "transformative," "in this fast-paced world."
  • Industry clichés specific to your niche: for coaches — "elevate," "abundance," "high-vibe"; for B2B — "drive ROI," "move the needle," "double down."
  • Hedging language: "may," "could," "in many cases," "tends to" (when used to soften every claim).
  • Personal idiosyncrasies: phrases you've spotted yourself overusing and want to retire.

Add to this list quarterly as you spot AI defaults creeping into output. Treat it as a living document.

SECTION 4

Tone by context

The same voice should shift between formats — your email voice isn't your sales page voice isn't your social voice. Specify the shifts.

For each format the AI will write in, note: average warmth (formal to warm), urgency (none to high), specificity expected (high to very high), and any format-specific rules.

Example structure:

Email to clients: Warmer than other formats. First-person, conversational. Often opens with a personal observation before the business point. Closes with a question.

LinkedIn post: Direct. Hook in first line. 150-300 words. Specific numbers in body. No external links unless absolutely necessary.

Sales page: Sharper, more direct than other formats. Names objections explicitly. Specific outcomes with specific timelines.

Blog post: Most analytical of the formats. 1,500-2,500 words. Frameworks named explicitly. Section headings descriptive, not clever.

SECTION 5

Signature moves

The 3-5 distinctive habits that make the writing recognisably this writer. The fingerprints.

Three to five maximum — more than five dilutes the instruction and the AI starts treating signature moves as suggestions rather than rules.

Examples of common signature moves:

  • Always opens with the reader's pain before naming the solution.
  • Names a specific number in every post (or every section).
  • Uses contrast pairs: "Not A. B." or "It's not X. It's Y."
  • Closes with a one-line provocation rather than a CTA.
  • Frequent rhetorical questions, especially mid-paragraph.
  • Cites a specific year + claim ("In 2024, X happened").
  • Uses parentheticals frequently (like this one).
  • Always shows working before stating the conclusion.

To identify your own: paste 10-20 samples into ChatGPT and ask "What are the three to five distinctive habits in this writing that an imitator would need to copy?" The answer is your signature moves.

A real voice prompt example (~600 words)

VOICE ESSENCE Talks like a smart friend who's done this before. Direct, specific, slightly opinionated. Uses short sentences for emphasis and longer ones for explanation. Names the reader's pain before naming the solution. Refuses to play the latest trend until the data justifies it. MECHANICAL RULES - Sentence length: 8-22 words. Frequent short sentences for emphasis. Longer sentences for explanation. - Paragraph length: 1-3 sentences. One-sentence paragraphs allowed and encouraged. - Contractions: always. - Punctuation: em dashes yes (frequent). Semicolons no. Ellipses rarely. Avoid double-spaces between sentences. - Sentence openers: starting with "And" or "But" is fine. Rhetorical questions allowed mid-paragraph. - Lists: only when the content is genuinely list-shaped. Default to inline prose. - Voice: active voice 90%+ of the time. - Word count: LinkedIn posts 150-300 words; blog intros under 150; FAQs 60-80 words. BANNED WORDS "leverage," "cutting-edge," "thought leader," "best-in-class," "synergy," "delve into," "navigate" (in marketing context), "streamline," "robust," "seamlessly," "tapestry," "elevate," "transformative," "in this fast-paced world," "drive ROI," "move the needle," "double down" (overused), "may," "could" (when used as hedges), "in many cases," "tends to," "unlock," "utilise" (use "use"), "synergistic." TONE BY CONTEXT Email: warmer than other formats. First person, conversational. Opens with a personal observation. Closes with a question. LinkedIn post: direct. Hook in first line. Specific numbers in body. 150-300 words. Sales page: sharper. Names objections explicitly. Specific outcomes with specific timelines. Blog post: most analytical. Frameworks named explicitly. Descriptive section headings, not clever. Comments on others' posts: 1-3 sentences. Always include something only this person would say. Never "great post!" SIGNATURE MOVES 1. Names a specific number in every post (e.g., "47 builds shipped," "£497 founder pricing," "70-85% match"). 2. Uses contrast pairs: "Not A. B." or "It's not X. It's Y." 3. Names the reader's pain before naming the solution. 4. Closes with a one-line provocation rather than a CTA when posting on social. 5. Refuses to commit to a stance without showing the working.

The testing protocol

Run 5-10 test prompts after building the voice prompt. Each tests a different content task:

  1. "Write a LinkedIn post about [common topic in your work]."
  2. "Draft a 200-word email pitching [your service]."
  3. "Rewrite this paragraph in my voice: [paste a generic AI paragraph]."
  4. "Give me 5 hooks for a post about [topic]."
  5. "Write the intro of a sales page about [product]."
  6. "Draft 3 LinkedIn comments on this post: [paste a post]."
  7. "Write a 60-word FAQ answer to: [question]."
  8. "Open a blog post about [topic] with the first 150 words."

Read each output. Apply this test sentence by sentence: would I actually write this?

If the answer is no for any sentence, identify which voice prompt rule failed and tighten it. Common failure modes:

Three iteration rounds is typical to get to 80%+ first-draft voice match. After that, the voice prompt is stable. Quarterly review thereafter is enough maintenance.

The four mistakes that kill voice prompts

Mistake 1: Vagueness. "Friendly tone" produces default output. Specificity ("12-18 word sentences, always contractions, never semicolons") produces consistent output.

Mistake 2: Skipping the iteration step. The first version of the voice prompt is wrong. The second version is closer. The third version usually works. Don't expect first-build output to land — iterate.

Mistake 3: Never updating it. Voices evolve. Your business evolves. A voice prompt from 6 months ago might miss recent shifts. Quarterly review (15 minutes) is enough.

Mistake 4: Only using it for long-form. The voice prompt is most powerful when applied to everything — emails, captions, comments, ad copy. The more consistently you use it, the better the AI gets at approximating your voice on first draft.

Where to go from here

Skip the 3-hour build — get the voice prompt built for you

DFY Voice System uses The Voice Build methodology — the same five-section structure above — executed for you in 2-3 working days. Voice prompt + custom GPT + hook library. £497 founder pricing.

See The Voice Build

Frequently Asked Questions

What is a voice prompt?

A 500-800 word reference document describing how a person writes, fed into AI before any task to produce on-voice output.

How long should a voice prompt be?

500-800 words. Shorter is too vague; longer eats context window.

What's the most common mistake when building one?

Being vague. "Friendly and professional" produces default output. Specificity ("12-18 word sentences, always contractions") produces consistent output.

How do I test if my voice prompt works?

Run 5-10 test prompts across different content tasks. Read each output sentence by sentence. Ask "would I write this?" If no, tighten the failed rule.

Can I copy someone else's voice prompt?

Copy the structure (five sections), not the content. Voice prompts are by definition specific to one writer.