TL;DR
Feed an AI model 5 sentences you love, 3 of your own you wince at, and 3 competitor sentences you want to sound nothing like. Ask for a 1-page guide: four voice sliders (each picks a side, no “balance”), three voice principles with do/do-NOT examples, and two 10-word vocabulary lists. Then store the finished guide where your team writes — a Claude Project, a ChatGPT Project, or a Gemini Gem — so every draft starts on-brand. The whole thing takes an afternoon. The 27-page brand book takes a quarter and gets read by nobody.
The task
You read the landing page draft Monday morning, the launch email Monday afternoon, and the help-center FAQ on Tuesday, and they sound like three different companies. Three people wrote them and you have no document to point them to. The 27-page brand book your previous agency wrote is collecting dust because nobody reads it before writing a tweet. You want a 1-page tone guide tight enough that a new copywriter writes “on brand” by the end of week one, and tight enough that the existing team actually checks it before publishing.
This matters more in 2026 than it did three years ago: when half your drafts are AI-assisted, an unguided model defaults to bland-corporate, and a recent industry survey found 78% of consumers say they prefer brands that “sound human.” A tone guide is the artifact that keeps both your writers and your AI tools from drifting into generic.
Where AI helps, and where it does not
AI is excellent at distillation. Give it 10 of your real sentences plus 3 you flagged as “not us,” and the model reverse-engineers a voice profile that explains the difference. It is also good at producing the do/do-not vocabulary lists that catch most off-brand drift. Where AI fails: deciding where your brand should sit on the tone spectrum if you have not anchored it. Without 5 sample sentences you genuinely love, the model defaults to “professional but approachable,” the most useless tone guide ever written. Anchor with real samples from your own writing; the model finds the pattern, you make the judgment call.
A common failure mode: the model outputs a guide full of “X but also Y” balances (“casual but professional,” “confident but humble,” “playful but trustworthy”). These read smart and produce zero behavior change, because writers can justify any sentence as “the balance point.” Tell the model up front: every slider must pick a side, and every “but” must die.
What to feed the AI
- 5 sample sentences you wrote that feel perfectly on-brand (emails, landing pages, social posts, FAQ answers; mix the sources)
- 3 sample sentences from your own past content that feel not us, the ones that make you wince. These are gold.
- 3 competitor sentences in your category that you actively want to sound different from
- Your audience in one sentence: who they are, what they care about, what they do for a living
- The one action you want them to take after reading you: sign up, click, share, buy, refer
- The 3 strongest emotions you want a reader to feel: calm, sharp, witty, reassured, energized
- The platforms you publish on. A tweet and a help-center article demand different lengths, so the slider can shift slightly between them.
- The 2-3 phrases you are tired of seeing in your category (“revolutionary,” “next-gen,” “delight”)
Which AI tool to use (June 2026)
The prompt below works in any frontier chat model. The real decision is where the finished guide lives so every future draft starts on-brand instead of re-pasting the guide each time. All three major platforms now offer a persistent workspace for exactly this:
| Tool | How you store the guide | Tier needed | Note |
|---|---|---|---|
| Claude Projects | Paste guide into Project custom instructions; Sonnet 4.6 / Opus 4.7 | Pro $20/mo | Strongest instruction-following of the three; also has “Styles” for tone |
| ChatGPT Projects / custom instructions | Project instructions, or a Custom GPT; GPT-5.5 | Free tier works; Plus $20/mo for limits | Most teams already have access |
| Gemini Gems | Build a Gem; Gemini 3.1 Pro | Free tier; better on Google AI Pro $19.99/mo | Gems link live to Google Drive, so the guide updates when your Doc does |
Prices and model versions as of June 2026. For pure distillation quality, Claude Opus 4.7 and GPT-5.5 are both strong; if your tone guide lives in a Google Doc your team already edits, a Gemini Gem keeps it in sync automatically. Whichever you pick, paste the 1-page guide into the workspace instructions once, and your future drafts stop drifting.
Copy-ready prompt
Distill a 1-page brand tone guide from the samples below.
On-brand samples (the voice we want): [paste 5 sentences]
Off-brand samples (our own past, the voice we want to avoid): [paste 3 sentences]
Competitor sentences we want to sound DIFFERENT from: [paste 3 sentences]
Audience: [one sentence]
Action we want readers to take: [one specific action]
Top 3 emotions we want readers to feel: [list]
Platforms we publish on: [list]
Category cliches we are tired of: [list]
Return:
1) Tone position on 4 sliders, each scored 0 to 10 with a one-sentence
rationale. The 4 sliders: formal/casual, serious/playful,
warm/distant, simple/sophisticated. Pick a side. Do not write "balance."
2) 3 voice principles. Each gets a one-sentence statement plus 1 do-example
and 1 do-NOT example pulled from my samples.
3) A 10-word "use these" vocabulary list (verbs and nouns we should prefer).
4) A 10-word "do-not use" list, including the category cliches I gave you.
5) Edit-this-sentence exercise: take the generic line
"We're excited to announce our newest features that empower users to do more."
and rewrite it twice, on-brand, with a one-sentence justification each.
Format: 1 page. No "but," no "balance," no "professional yet approachable."
If a slider rationale contains a "but," you have not finished thinking.Shorter variant: voice audit on existing copy
Audit the copy below against this brand tone guide: [paste 1-page guide].
Copy: [paste 5-10 sentences].
For each sentence: rate 0-10 on each of the 4 voice sliders, and flag any
phrase that violates the "do not use" list.
Then rewrite the 3 weakest sentences on-brand, and briefly explain what
made each one off-brand.Sample output
A useful slider rationale (no “but”): “Casual: 7/10. We use contractions (‘we’ll,’ ‘don’t’), short sentences, and verbs that do real work (‘ship,’ ‘pull,’ ‘drop’ over ‘release,’ ‘extract,’ ‘announce’). We never use slang we’d be embarrassed to read in 3 years.”
A useful voice principle: “Specific over impressive. Do: ‘cut your onboarding email queue by 80% in week one.’ Do NOT: ‘transform your customer engagement with AI-powered automation.’ We earn trust by saying what actually happens, not by sounding sophisticated.”
A useful do-not list: “Avoid: revolutionary, next-gen, robust, leverage (as verb), delight, empower, supercharge, unlock, seamlessly, world-class, innovative.” Each gets replaced by the underlying specific thing it was trying to claim.
A useful rewrite of the generic line: Original: “We’re excited to announce our newest features that empower users to do more.” Rewrite 1: “Three new things shipped today. Here’s the one we’re most excited about.” Rewrite 2: “You can now [specific feature]. The full changelog is here.” Both rewrites refuse the empty verbs (“excited to announce,” “empower”) and lead with the specific.
How to refine
- If the guide reads as “casual but professional” (useless): “Kill every ‘but.’ Each slider picks a side. If two sliders feel contradictory in practice, write 5 test sentences that satisfy both and rate them.”
- If the do-not list misses obvious offenders: “Add every verb I never want to see (‘leverage,’ ‘empower,’ ‘unlock,’ ‘supercharge,’ ‘delight’). These are the words my team writes when they have nothing specific to say.”
- If the voice principles are too abstract: “Each principle needs a do-example AND a do-NOT example pulled from the sentences I gave you. If the example does not come from my samples, the principle is too generic.”
- If new hires still sound off-brand: “Add a 5-sentence ‘baseline drift test’: five representative sentences a new writer rewrites in week 1; I evaluate against the guide and flag where they drift.”
- If the guide grows beyond one page: “Cut to one page. Anything that survives is signal. Vocabulary lists belong; meta-commentary about how to use the guide does not.”
Common mistakes
- Adopting a voice that does not match how you actually speak. The guide gets ignored after week 2 because writing in a borrowed voice is exhausting. Pick the voice that is a sharpened version of how the founder or strongest writer actually writes.
- Skipping the off-brand list. Half the value of a tone guide is what NOT to do; without examples of off-brand, writers cannot self-correct.
- Letting AI pick the voice for you. The model finds patterns in your samples; it does not decide the brand. If you have not made the call, the guide will be generic.
- Writing a 27-page brand book. Nobody reads it, nobody references it, the document might as well not exist. One page or it does not get used.
- No specific do-not vocabulary. Abstract principles like “be specific” are unenforceable; “never use these 10 words” is.
- Updating quarterly. Voice drift is real, but quarterly updates train your team to assume the rules will change. Annual updates with a clear changelog note are better.
- No example of an edited sentence. Without before/after, the guide is theory; the rewrite exercise is what makes the principles concrete.
- One slider position for all platforms. Your help-center FAQ and your launch tweet can sit at different points; the guide should call out the platform-specific shift.
FAQ
- How long should the tone guide be? One page. Two pages and your team skips it. If your principles cannot fit on one page, they are not tight enough; sharpen, do not expand.
- How often should I update it? Annually. Voice does shift, usually following audience growth or product evolution, but quarterly updates train the team that rules are unstable, and stability is what makes the guide usable.
- Should each platform have its own variant? One master guide plus a small “platform notes” appendix is the right shape. The voice principles stay; the slider position can shift slightly (casual a notch higher on Twitter, simple a notch higher in product copy).
- Which AI model writes the best guide? As of June 2026, Claude Opus 4.7 and GPT-5.5 both distill voice profiles well; Gemini 3.1 Pro is close behind. The bigger choice is storage: keep the finished guide in a Claude Project, ChatGPT Project, or Gemini Gem so every future draft inherits it without re-pasting.
- What about voice for technical content (API docs, error messages)? Voice principles apply; vocabulary changes. Tighter, more precise, fewer adjectives. Keep a separate vocabulary list for technical writing but the same 4 sliders.
- How do I get the team to actually use it? Two moves: (1) include the do-not list in your weekly editorial review so off-brand words get flagged in PRs and drafts, and (2) end every new hire’s first week with a 5-sentence rewrite exercise against the guide. The guide enters muscle memory through that exercise, not through being “read.”