By 
Shreyas M V · Published · 14 min read
Brand Voice Documents for AI Writers: What to Include, With a Template
A brand voice guide for AI writers is a document that describes how your company sounds in enough detail that a writer who cannot ask you questions can still match it: personality, tone by situation, rhythm, vocabulary, banned phrases and worked examples. This page gives a ten-section template and a drift test that checks whether AI drafts still match the document over time. The template draws on two published guides. The drift test is Kanso's own method, and no measured result backs it yet.
- A voice guide works when it covers four facets (personality, tone, rhythm and vocabulary) and shows each with examples, not when it lists six adjectives.
- The template below has ten sections, and the table that introduces them says which guide each idea comes from.
- The drift test runs a fixed set of briefs through your AI tool, scores the unedited drafts from 0 to 2 per facet, counts banned phrases and sentence length, and compares each run with a baseline.
- Both guides read for this page are practitioner writing. Nothing here measures how much a voice document improves AI output.
What is a brand voice guide for AI writers?
A brand voice guide for AI writers is a written description of how your company sounds, detailed enough that a writer who cannot ask you a question can still match it.
The same document can serve freelancers and AI tools, but the AI case is stricter. A freelancer can message you with a doubt. A model works only from the page in front of it.
The common failure is a guide that is too thin. Lauren Pope, who writes about voice from her consulting work, asks whether your guidelines are six adjectives with a line of explanation each, or whether they exist only in one person's head. She says that in years of consulting she could count on one hand the organisations that gave her a brilliant voice guide to work from.
What is the difference between brand voice and tone?
Voice is the constant character of how a brand writes, and tone is how that voice shifts for a given audience and situation. Pope uses a coffee analogy: personality is the coffee, and tone is the milk or water you add to suit the order.
The Branded makes the same split in its own terms. Voice is the core persona, and tone is the mood and manner that change by channel or audience.
This matters more for AI writers than for people. A document that defines only voice gives a model one register for every task, so a refund apology and a launch announcement come out sounding alike. That is Kanso's analysis, not a finding from either guide.
Which sections should a brand voice guide include?
A brand voice guide should include ten sections: owner and version, reader, personality, tone by situation, rhythm, vocabulary, banned phrases, terminology, claims rules and examples. The table shows what goes in each and where the idea comes from.
| # | Section | What to write | Origin |
|---|---|---|---|
| 1 | Owner, version and date | Name the one person who approves changes, the version number and the next review date. | Kanso suggestion |
| 2 | Reader | Say who you write for and what they want from you. | Pope (tone starts from the audience); Branded (research your audience) |
| 3 | Personality | A short character description, then three to five traits, each with what it means and what it does not. | Pope |
| 4 | Tone by situation | One card per situation: the -ing verb, who it is for, when to use it, which traits to turn up or down. | Pope |
| 5 | Rhythm | Sentence and paragraph length ranges, punctuation habits, and when short sentences are used. | Pope names rhythm; the numbers are a Kanso suggestion |
| 6 | Vocabulary | Words you use, and whether the register is plain or specialist for each reader. | Pope |
| 7 | Banned phrases | Phrases that are off brand, each paired with what to write instead. | Pope |
| 8 | Terminology and stance | Preferred names for people and things, and inclusive-language rules. | Pope; Branded (inclusive language) |
| 9 | Claims rules | What may be stated as fact, what needs a source, and a ban on invented numbers, customers and quotes. | Branded (be accurate); the AI-specific wording is a Kanso suggestion |
| 10 | Examples | An on-voice and an off-voice version of the same line for every tone card, plus two full sample paragraphs. | Pope (examples); Branded (sample responses) |
How the table was put together: Pope's four facets (personality, tone, rhythm, vocabulary) supply sections 3 to 6, and her advice on off-brand words and terminology supplies 7 and 8. The Branded supplies the reader, accuracy and inclusive-language points. Sections 1 and 9, and every numeric limit, are Kanso's additions for AI use.
Section 2 is where customer language earns its place. If you do not yet know how your buyers describe their problem, the method for turning reviews and threads into copy shows how to collect the phrases first.
How do you write personality and tone so an AI can use them?
Write personality as a short character description with examples, and write tone as one card per situation, because a list of adjectives gives a writer nothing to imitate. Pope's advice is to go beyond adjectives by writing a character description that helps people assume the personality, then adding copy examples of each trait in action.
Her examples show how different a good answer can be. She describes BrewDog as passionate, independent, punk and a little bombastic, and GOV.UK as neutral and fair. Each fits what the organisation is for, which is her test of authenticity.
For tone, Pope suggests naming each one with an -ing verb, such as informing, warning, thanking, supporting or championing. For each tone, describe who it is for, when to use it, which personality traits to turn up or down, and show it in action. The table applies that structure to a fictional payroll company called Acme Payroll.
| Situation | Tone | Turn up | Turn down | On-voice line |
|---|---|---|---|---|
| Announcing a feature | Announcing | Plain, specific | Enthusiasm | You can now export payslips as one file. It takes one click from the Reports page. |
| Apologising for a late payroll file | Owning | Directness, calm | Humour, brand personality | Your June payroll file was two days late. That was our error, and here is what we changed so it does not repeat. |
| Explaining how overtime is calculated | Explaining | Precision, patience | Slang, shortcuts | Overtime is calculated per week from the hours in your approved timesheet. Here is the formula and one worked example. |
Brand archetypes are a third route to personality. The Branded recommends choosing an archetype, defining traits that express it and letting it shape every campaign. Its page says there are twelve archetypes but names eleven (Sage, Innocent, Explorer, Hero, Outlaw, Creator, Jester, Lover, Caregiver, Magician and Ruler), so check any archetype list you copy against its source.
How do you document rhythm and vocabulary?
Document rhythm as numbers and habits a reader can check, and vocabulary as two lists: words you use and words you never use. Pope defines rhythm as the pace and pattern of a voice, and notes that short sentences are impactful, alliteration can bring playfulness and varied sentence length keeps a voice interesting.
She points to Oatly as a voice with an offbeat rhythm: line breaks in odd places and unexpected punctuation that make you work a little to read it. A rhythm section for such a brand would need to say so explicitly, or a model will smooth it out.
Pick ranges you can count. For example, write that sentences average 12 to 18 words and that no paragraph runs past three sentences. Those numbers are Kanso's suggestion, chosen because they can be measured in the drift test below.
On vocabulary, Pope says the words you use might be simple or complex, narrow or wide, plain or specialist, and may shift with tone and audience. She also recommends a terminology section where your organisation takes a stand. In her example for a fostering and adoption charity, it covered how to describe children and their birth and adoptive parents.
How do you write a banned-phrases list that works?
A banned-phrases list works when every entry is a specific phrase paired with what to write instead, and when it is short enough to check by search. Pope calls a list of words that are not on brand, with better alternatives, a practical tool for writers.
| Banned phrase | Why it is off voice | Write instead |
|---|---|---|
| game-changer | Says nothing a reader can check | Say what changed, with the number or the step removed |
| seamless | Hides the step the customer still does | Name the step that is gone |
| unlock | Treats a feature as a locked door | let you, or start |
| in today's fast-paced world | A warm-up sentence that carries no fact | Delete it and open with the point |
| we're thrilled to announce | Announces the feeling, not the news | We released, followed by the thing |
| leverage | Corporate register for a plain verb | use |
Start from your own editing history. The first entries should be the phrases your editors already delete from drafts, because those are the ones you will keep paying to remove.
A banned phrase is also the cheapest drift signal, because finding it needs a text search and no judgement. Its limit is that banning one phrase does not stop a near synonym, so the drift test below scores the facets as well as counting phrases.
What examples should a brand voice guide contain?
A guide should contain an on-voice and an off-voice version of the same line for every tone card, plus two full sample paragraphs. Pope says copy examples help make personality traits and tones tangible, and The Branded gives sample customer responses, such as thanking a buyer for an order and explaining a complex issue in plain steps.
The off-voice version is Kanso's addition. A model told what to sound like still has to guess what to avoid, and an off-voice example answers that directly.
One observation about the sources shows why examples outrank rules. The Branded's tip list advises clear writing with short sentences, simple words and no jargon, yet the page itself is written in a heavily idiomatic register. That is an observation about one page, not a verdict on its advice, but it shows that a rule list and a writing sample can point in different directions.
What does a copy-ready brand voice template look like?
A copy-ready template is the ten sections as fill-in headings in one document, with the rules a writer must never break kept short and near the top of each section. Copy the list below into a document and replace each bracket.
- Owner, version, date. Owner: [name]. Version: [1.0]. Last reviewed: [date]. Next review: [date].
- Reader. We write for [role] who wants [outcome] and worries about [concern]. They already know [topic] and do not know [topic].
- Personality. If our voice were a person, they would be [two or three sentences]. Traits: [trait] means [meaning] and does not mean [limit]. Give three to five traits, with one on-voice line each.
- Tone cards. For each situation: Situation: [name]. We are [-ing verb]. For [reader]. Use when [trigger]. Turn up [trait]. Turn down [trait]. On-voice line: [line]. Off-voice line: [line].
- Rhythm. Sentences average [range] words. Paragraphs run [number] sentences at most. Punctuation: [habits]. We use a short sentence when [purpose].
- Vocabulary. We use [words]. Our register is [plain or specialist] for [reader]. We spell [term] as [spelling].
- Banned phrases. [Phrase] becomes [replacement]. Keep the list to about fifteen entries.
- Terminology and stance. We call [thing] a [term], never [other term]. Inclusive-language rules: [rules].
- Claims rules. Every figure comes from the brief or a named source. Never invent numbers, customers, quotes or dates. When a fact is missing, write [fact needed] and stop.
- Examples. One on-voice and one off-voice line per tone card. Two sample paragraphs of [length], marked as approved.
Where you paste the finished document depends on the tool. Most AI writing tools have a place for standing instructions, but the field names differ, so check yours and put the document there rather than repeating it in every request.
What changes when the reader of your guide is an AI writer?
Three things change: the document must stand alone, its rules must be checkable, and its examples carry more weight than its descriptions. These are Kanso's working assumptions. None of the sources read for this page measures how AI tools respond to voice documents.
- Stand alone. Remove phrases such as “you know our tone”. Every term and trait should be explained inside the document.
- Checkable rules. A rule such as “sentences average 12 to 18 words” can be counted. “Be conversational” cannot.
- Facts in the brief, voice in the document. The voice document says how to write. The brief says what to say, including every figure, so the claims rule has nothing to fill from imagination.
Forum threads on AI and branding raise a worry that AI tools push every brand toward the same sound. The research for this page holds no measurement of that, so treat it as an open concern. Pope's requirement that a voice be differentiated gives you something to write against: say what you never sound like, and include an off-voice example in a generic competitor's style.
An approval step still belongs after the draft, whatever the document says. The guide to approval gates in AI marketing workflows covers where to put it.
How do you test for brand voice drift?
Test for drift by running the same fixed set of briefs through your AI tool on a schedule, scoring the unedited outputs against your document, and comparing each score with a baseline run. Voice drift is the gap that opens between what the document says and what the drafts sound like, whether the cause is a change in the tool, the model, the prompt or the document.
This is Kanso's method. It uses Pope's four facets as the scoring criteria and the consistency attribute from The Branded as the goal. Neither guide describes a way to test for drift in the text extracted for this page, and Kanso has not measured how well this test detects it.
- Freeze the document. Give it a version number and a date. If you edit it mid-test, you will not know what caused a change.
- Write six briefs. Cover different situations from your tone cards, such as an announcement, an apology and an explanation, and keep the wording fixed for every run. The format in the content brief template works, and each brief should contain every fact the draft needs.
- Generate a baseline. Use the tool and settings you use in production, with the document loaded. Save the outputs unedited and record the tool, the model name and the date.
- Count first. Run the mechanical checks in the second table below. They need no judgement, so they cannot be argued with.
- Score the facets. Rate each output from 0 to 2 on personality, tone, rhythm and vocabulary using the first table below. Six briefs at four facets and a top score of 2 give a maximum of 48.
- Run a blind check. Mix two pieces your team approved into the set, remove labels, and ask a second reader to mark each piece “ours” or “not ours”. Treat the result as a prompt for discussion, not a measurement.
- Set thresholds before the next run. Kanso's starting values: flag drift if the total falls by 4 or more points (4 of 48 is about 8%), if any facet falls by 3 or more of 12, if any banned phrase appears, if average sentence length leaves the range, or if any invented fact is found.
- Re-run on triggers. Re-run after any change to the document, the tool or the model, and on a calendar interval you can keep, such as monthly.
| Facet | Score 0 | Score 1 | Score 2 |
|---|---|---|---|
| Personality | Reads like no one in particular | Some listed traits show, others are missing | Every listed trait shows and none is contradicted |
| Tone | Wrong mood for the situation | Right mood, but ignores the card's turn-up and turn-down notes | Matches the tone card for that situation |
| Rhythm | Outside the document's ranges throughout | Mostly in range, with runs outside it | In range, using the stated habits |
| Vocabulary | Ignores the document's terminology or word choices | A few lapses | Uses the document's terms and no off-list words |
| Check | How to count it | Pass rule |
|---|---|---|
| Banned phrases | Search every output for every entry on the list | 0 hits |
| Average sentence length | Total words divided by total sentences, across all outputs | Inside the range written in section 5 |
| Longest paragraph | Count the sentences in the longest paragraph | At or under the document's limit |
| Required terms | Search for the preferred terms where a brief calls for them | Present, with no rival term |
| Invented facts | Mark any number, customer, quote or date that is not in the brief | 0 found |
What does a scored drift test look like?
A scored drift test is a table of baseline and re-run scores for each facet, with the counted checks beside it. Every number in this section is invented to show the arithmetic and is not a measured result from Kanso or anyone else.
| Measure | Maximum or rule | Baseline | Re-run | Change |
|---|---|---|---|---|
| Personality | 12 | 11 | 10 | -1 |
| Tone | 12 | 10 | 7 | -3 |
| Rhythm | 12 | 10 | 6 | -4 |
| Vocabulary | 12 | 10 | 9 | -1 |
| Total score | 48 | 41 | 32 | -9 |
| Banned phrase hits | 0 allowed | 0 | 3 | +3 |
| Average words per sentence (target 12 to 18) | - | 15 | 23 | +8 |
The arithmetic: the baseline total is 11 + 10 + 10 + 10 = 41, and the re-run total is 10 + 7 + 6 + 9 = 32, a fall of 9. Average sentence length is words divided by sentences: 1,260 / 84 = 15 at baseline, and 1,380 / 60 = 23 on the re-run.
Against the thresholds above, this re-run fails on four counts. The total fell by 9 against a limit of 4, and tone (3) and rhythm (4) each fell by at least the facet limit of 3. Three banned phrases appeared, and the sentence average of 23 is outside the 12 to 18 range.
Reading it by facet points to the cause. Rhythm and sentence length fell together, which suggests the rhythm section is not holding, and tone fell where a situation had no card or a weak one. Vocabulary moved by one point despite three banned phrases, which shows why the counted check sits beside the rubric.
What do you change in the document when the test fails?
Change the document, not the individual draft, because a hand-edited draft hides the gap that will reappear in the next one. The table maps common symptoms to the likely gap and the edit to make. It is Kanso's working guide, built from the structure of the template, not a tested diagnosis.
| Symptom | Likely gap | Edit to try |
|---|---|---|
| Correct but flat | Personality is a list of adjectives | Add a character description and two on-voice lines per trait |
| Right words, wrong mood in a hard moment | No tone card for that situation | Add a card with the -ing verb, the trigger and an on-voice example |
| Sentences creep longer | Rhythm section has no numbers | Write a sentence range, a paragraph limit and one short-sentence rule |
| Banned phrases return | List is long, buried or has no replacements | Shorten it, move it up, add a replacement for each entry |
| Invented statistics or customers | Claims rule is missing, or the brief does not supply the facts | Add the rule, and put the real figures into the brief |
| Sounds like every other brand | The document never says what you do not sound like | Add an off-voice example in a generic competitor's style |
After the edit, bump the version number and run the same six briefs again. If the re-run scores within the threshold of the baseline, make it the new baseline and note the date. In the invented example, a re-run total of 40 would be 1 point below the baseline of 41 and would pass.
freeze the document → generate with fixed briefs → count → score → compare with baseline → edit the document → repeat
What is documented on this page and what is Kanso's analysis?
The documented points come from two published guides read on 28 October 2026, and everything else is Kanso's analysis. Neither guide is a study, and the research package held no primary source and no measured data on AI output.
- Documented: Pope names four facets (personality, tone, rhythm, vocabulary) and four qualities of a good voice (authentic, reflective, differentiated, readable).
- Documented: Pope advises -ing verbs for tones, character descriptions and copy examples for personality, and a list of off-brand words with alternatives.
- Documented: The Branded separates voice from tone and offers an archetype exercise, and its page says twelve archetypes while naming eleven.
- Kanso's analysis: The ten-section order, the numeric limits, the claims rule wording and the off-voice example requirement.
- Kanso's analysis: The three AI-specific assumptions, the drift test, the rubric, the thresholds and the symptom table. None has been run on real drafts for this page.
- Invented for illustration: Acme Payroll, its tone cards and banned phrases, and every score in the worked example.
Where does Kanso fit into writing and testing a voice document?
Kanso builds a Brand Voice document from your website as one of the company documents its agents draft from, but it does not run the drift test in this article. As of October 2026, Kanso's homepage FAQ says it crawls your site and public content, then generates product, brand voice, competitor and content strategy documents that its agents work from.
The Writer Agent turns Search Console opportunities into long-form article drafts in your brand voice, for you to review and publish yourself. The LinkedIn Agent and X Agent draft posts from your marketing context, which you edit and publish when you are ready. Kanso does not post for you.
Read the generated document against the ten sections above, then run the drift test on the drafts the agents produce. The founder voice use case describes the LinkedIn and X workflow,, and the pricing page has the plan details.
Brand voice guide: frequently asked questions
What is a brand voice guide?
A brand voice guide is a written description of how a company sounds, covering its personality, tone, rhythm and vocabulary, with examples a writer can imitate. Lauren Pope, writing on Medium, names those four facets and says many organisations have only a list of adjectives, or no guide at all because the voice sits in one person's head.
What is a tone of voice guide?
A tone of voice guide is the part of a voice guide that sets how the voice shifts for different audiences and situations. Pope describes tone as the feelings or moods a voice can convey, while personality stays constant, and suggests naming each tone with an -ing verb such as informing, warning or thanking.
Why is a brand voice important?
A consistent voice helps readers recognise who is speaking, and The Branded argues it builds trust and loyalty. That is the agency's claim and its page gives no study behind it. Pope adds that voice also affects product and user experience, not only marketing.
How do I follow a brand voice if I am not given any guidelines?
Ask for three pieces the owner loves and three they rejected, then read them for the four facets: personality, tone, rhythm and vocabulary. The Branded's outline includes an Inspect Your Current Content step for the same reason. Write what you find as a draft guide and have the owner approve it.
How can I make an AI tool follow our brand tone of voice consistently?
Give the tool a self-contained voice document with examples, then test it on a fixed set of briefs on a schedule. The document should state checkable rules such as sentence ranges and banned phrases, because a model cannot ask what you meant. The drift test on this page is a method for the second half; it is Kanso's suggestion and has not been validated.
Can AI help shape a brand's voice, or does it make everything sound the same?
Both outcomes are possible, and nothing in the research for this page measures which is more common. Pope lists differentiation as a requirement of a good voice, so write down what you do not sound like and include an off-voice example. Then use the drift test to catch the pull toward generic phrasing.
Should I use the twelve brand archetypes to define a voice?
Use archetypes as a brainstorming prompt for personality, not as the finished document. The Branded recommends picking an archetype and writing traits that express it, but its page says twelve archetypes and names eleven. An archetype also says little about rhythm or vocabulary, which a voice guide still needs.
What if the founder who set the voice has moved on?
Rebuild the voice from the founder's approved work and put it in a document with a named owner, a version and a review date. Pick ten approved pieces, read them for the four facets, and write the findings as sections 3 to 7 of the template. This is a Kanso suggestion, not a published procedure.
Are there good exercises for developing a brand voice?
The two guides read for this page offer three exercises: write a character description for the personality, name each tone with an -ing verb, and list off-brand words with better alternatives. The research does not supply a book list, so none is recommended here. The Branded also offers an archetype exercise.
Where can I get a brand voice template?
The ten-section template on this page is free to copy. A Notion template listing for a brand and voice guide ranked first when the research was run on 28 October 2026, but only 13 words of it could be extracted, so its contents are not covered here.
How do I test whether AI writing still matches our brand voice?
Run the same six briefs on a schedule, score the unedited outputs against your document, and compare the score with a baseline run. Score four facets from 0 to 2, count banned phrases and sentence length, and flag a drop of 4 or more points out of 48. The thresholds are starting values chosen by Kanso, not tested ones.
Sources
- The anatomy of a brilliant brand voice, Lauren Pope, Medium
- Step-by-Step Brand Voice Guidelines + Framework, The Branded (brandedagency.com), page dated 29 March 2025
- Writer Agent, Kanso
- Kanso homepage, frequently asked questions, Kanso
Note: Facts about brand voice are attributed to two published guides, read on 28 October 2026. Neither is a study with a stated method, and no primary source was available for this topic. The drift test, its thresholds and the template ordering are Kanso's own suggestions and have not been validated. Every number in the worked example is invented. Kanso has no affiliation with the authors of the guides cited.
AI workflows