How to Build a Brand Voice Guide That AI Can Follow

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Build a Brand Voice Guide That AI Can Follow.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Build a Brand Voice Guide That AI Can Follow.

Collect 10 to 20 pieces of your best existing writing, have AI extract the patterns, then turn them into rules a model can follow: three or four voice traits, each with a "do" and "don't" example, a word list, banned phrases, sentence and formatting rules, and before-and-after rewrites. Keep it to two pages, load it into a Project, and test it.

Most brand voice guides fail with AI because they are written in adjectives. "Friendly, professional, approachable" describes nearly every business, so it tells a model nothing it wasn't going to do anyway. Models follow examples and concrete, checkable rules. They also follow them too literally: tell one to be "playful" and every post gets a pun. A guide that works for AI shows rather than describes, and sets limits as well as goals.

Follow me on Instagram@sagnikteaches

Step 1: gather 10 to 20 samples that sound like you at your best

The raw material for the guide is writing that already works. Aim for variety across channels, all of it written by you or your team, not by an agency or a previous AI tool.

Connect on LinkedInSagnik Bhattacharya
  • Emails that got warm replies or led to a sale.
  • Social posts customers commented on in their own words, not just liked.
  • Product descriptions you wrote yourself and are proud of.
  • Replies to customer questions or complaints that went well.
  • A transcript of you explaining your business out loud. Record a two-minute voice note answering "what do you make and why is it different?" and transcribe it; spoken language is often closer to your real voice than anything you've polished.

Leave out anything that feels off, anything old enough to describe products you no longer sell, and anything written in a hurry that you'd rewrite now. Twelve good samples beat thirty mixed ones.

Subscribe on YouTube@codingliquids

An illustrative set for a furniture maker: four customer emails about commissions, three Instagram captions about the workshop, three product descriptions (a dining table, a bench, a bookcase), one reply to a customer whose delivery was delayed, and a transcribed voice note about why she uses hand-cut joints.

Step 2: let AI find the patterns, then argue with it

Paste the samples into an assistant and ask it to describe what they have in common, with evidence:

Below are 12 pieces of writing from my business. Analyse the voice.
For each point, quote the exact phrases that show it.
1. Three to four traits that describe how we write (not generic
   words like "friendly" unless you can show what makes ours specific)
2. Typical sentence length and structure
3. Words and phrases we use often
4. Words we never use that businesses like ours often do
5. How we open and close messages
6. Anything inconsistent between pieces
Don't write any new copy yet.

[samples]

An illustrative extract of the analysis, and what the owner changed:

"Trait 1: Plain and material-focused. You name woods, joints and finishes specifically ('quarter-sawn oak', 'drawbored tenons') and rarely use adjectives about quality. Trait 2: Honest about trade-offs ('it will mark, and that's part of it')... Sentence length: mostly 8-15 words, with one long sentence per paragraph... You never use 'luxury', 'stunning', 'bespoke experience'... Inconsistency: two captions use exclamation marks; nothing else does."

What she changed: she agreed with the traits but rejected "rarely use adjectives about quality" as too absolute; she does say "properly made". The exclamation marks were from a helper's posts, and she ruled them out.

This step is where the guide becomes yours rather than the model's. Treat the analysis as a first draft to argue with, not a verdict.

Step 3: turn adjectives into rules a model can check

Every trait needs to become something a model (or a person) can check a draft against. A useful test: could two people disagree on whether a draft follows the rule? If yes, it needs to be more specific.

Vague traitTestable rule
FriendlyAddress the reader as "you"; use first names in emails; contractions are fine (we're, it's)
ProfessionalNo slang, no emoji in emails, every factual claim traceable to the product sheet
WarmOne sentence per piece that refers to the reader's own situation
HonestMention one limitation or trade-off in every product description
ExpertName the specific material, technique or measurement instead of saying "high quality"
ConciseSocial captions under 80 words; emails under 150; no more than one sentence over 25 words per paragraph
Playful (use with care)At most one light touch per piece; never in replies to complaints or delays

The last row matters. A limit on how often a trait appears is what stops AI turning a hint of humour into a running joke.

Step 4: write the guide itself

Below is the furniture maker's finished guide (illustrative), laid out in the format that assistants tend to follow best. Copy the structure and replace every line with your own.

VOICE GUIDE: [workshop name]      Version 1.2   Updated: {date}

WHO WE'RE TALKING TO
People furnishing a home they plan to stay in. They care about how
things are made and will pay for it, but distrust sales talk.

HOW WE SOUND (4 traits)
1. Material-first. Name the wood, joint and finish.
   Do: "Quarter-sawn white oak, drawbored tenons, hard wax oil."
   Don't: "Crafted from the finest premium materials."
2. Honest about trade-offs. One per product description.
   Do: "Oil finishes mark more easily than lacquer. They also
   repair at home in five minutes."
   Don't: "Built to be completely maintenance-free."
3. Plain and short. Mostly 8-15 word sentences; one longer one
   per paragraph is fine.
4. Warm, not gushing. One line about the reader's situation.
   Do: "A table for eight that folds down to four for weekdays."
   Don't: "Transform your dining experience!"

WORDS WE USE: properly made, joint, grain, finish, commission,
workshop, "it'll last", "we'd suggest"
WORDS WE NEVER USE: luxury, stunning, bespoke experience, elevate,
curated, timeless (unless quoting a customer), exclamation marks

FORMAT RULES
- House spelling: colour, organise, centre. Numbers as digits. Prices with $ and no "only".
- Emails: first name, under 150 words, sign off "Speak soon, [name]"
  (never on complaint replies: use "Thanks for your patience").
- Never promise lead times, prices or delivery dates not in the
  current price list.

BEFORE / AFTER
Before: "Our stunning handcrafted dining tables bring timeless
elegance to any home!"
After: "Each dining table is cut and jointed in our workshop from
oak we choose board by board. It takes about ten weeks."

WHEN UNSURE: shorter, plainer, more specific.

Notice what's in it: an audience line, four traits each with a do and don't, word lists in both directions, hard format rules, one before-and-after pair and a tiebreaker. Keep the full set of 12 samples in a separate file uploaded alongside it, so the model has more examples without the guide itself getting long.

The difference the guide makes is easiest to see on one task. The same request ("write an Instagram caption for our new walnut bench") with and without it, both illustrative:

Without the guide: "Introducing our stunning new walnut bench! Handcrafted with love and built to last a lifetime, it's the perfect statement piece to elevate any entryway. Shop now!"

With the guide: "New on the bench, and now a bench: black walnut, through-tenons left proud so you can see how it's held together. 120cm, seats three. Walnut darkens a little in sunlight over the first year, which we think is the best bit. In the showroom from Friday."

Everything in the second version is checkable, it names the material and the joint, it mentions a trade-off, and it contains one light touch ("now a bench") rather than a stream of them. That is the four traits working together.

Traits must also be specific to your business, not borrowed from a template. Compare the furniture maker's traits with those an e-commerce homeware brand might settle on after the same exercise: "sensory" (describe how things feel and sound: the weight of a mug, the rustle of linen), "practical" (every product description answers "will this work in my home?" with a size or care fact), and "calm" (no urgency language, no countdowns, no "don't miss out"). Both businesses sell home goods; the guides would produce clearly different copy, which is the point.

Step 5: add channel notes, not separate voices

The voice stays the same across channels; length, structure and what you lead with change. A short table at the end of the guide handles this. For the furniture maker:

ChannelLengthLead withExtra rule
InstagramUnder 80 wordsWhat's on the bench todayOne process detail per post; no hashtag walls (5 maximum)
Email newsletterUnder 250 wordsOne finished piece and its storyOne link, to the piece
Product page120-200 wordsWhat it is and what it's made ofDimensions and care in a list at the end
Enquiry repliesUnder 150 wordsAnswer their question firstPrice ranges only from the price list
Complaint or delay repliesUnder 120 wordsAcknowledge the problem plainlyNo humour, no sign-off flourish, a named next step and date

In a different business the rows change but the idea holds. A toy shop might add a rule for its Instagram row ("age range in every product post") and one for its email row ("never imply a toy is suitable below its packaging age"), because the voice guide is also where brand-critical facts get enforced.

Step 6: load it where your team actually drafts

  • ChatGPT Projects: paste the guide into the project's instructions and upload the samples file. OpenAI's help pages note that project instructions apply only inside that project and override your general custom instructions, so a "Marketing" project keeps the voice separate from everything else. Projects can be shared with a team, with chat or edit access.
  • Claude Projects: the same approach: guide in the project instructions, samples as project knowledge.
  • Gemini Gems (which become skills from November 2026): put the guide in the Gem's instructions and attach the samples.
  • Custom GPTs: don't start one. OpenAI is retiring custom GPTs, which stop running on 11 December 2026; use a Project instead.
  • Design and email tools with a brand voice field: Canva's Brand Voice, for example, caps the guidelines at 500 characters, so you need a compressed version.

A compressed version of the furniture maker's guide that fits a 500-character field:

Plain, material-first, honest. Name the wood, joint and finish
instead of praising quality. Mention one trade-off per product.
Mostly short sentences. Warm, not gushing: one line about the
reader's home. Spell colour, organise. Never use: luxury, stunning,
bespoke experience, elevate, curated, exclamation marks. Never
invent prices, lead times or delivery dates. When unsure: shorter,
plainer, more specific.

That comes to roughly 400 characters, leaving room for a line about the audience. Keep the master guide as a document you own, so moving tools means pasting it again rather than rewriting it from memory.

Step 7: test it blind before you trust it

Pick five real tasks you'll need this month: a product description, two social captions, an enquiry reply and a newsletter intro. Generate each with the guide loaded. Then mix the AI versions with pieces you wrote yourself and ask someone who knows the business to sort them.

Score each AI draft against a short rubric:

CheckPass looks likeIllustrative first-round result
Traits followedAll four traits present, none overdone4 of 5 passed; the newsletter intro gushed
Banned words absentNone of the list, and no obvious substitutesPassed, but "refined" appeared twice as a stand-in for "luxury"
Format rulesLengths, spelling, sign-offs correctEnquiry reply used the cheerful sign-off on a delay question
FactsNothing invented about price, time or materialsProduct description added "FSC-certified", which isn't true for this oak
Blind sortReader can't reliably pick the AI piecesReader picked 3 of 5 correctly

Each failure becomes an edit to the guide: "refined" joins the banned list, the delay rule moves higher up, "never state certifications" goes into the format rules. Run the test again with five new tasks. Two rounds usually get a guide to the point where drafts need light edits rather than rewrites.

Use the same guide to onboard people, not just AI

A guide written as testable rules works for new staff as well as models, which is a good sign it's specific enough. A sports equipment shop that wrote one for its AI tools found its Saturday staff used it too, when replying to messages on the shop's accounts. A ten-minute exercise for any new helper:

  1. Read the guide (three minutes).
  2. Rewrite one of the "don't" examples in the guide's voice (three minutes).
  3. Draft a reply to a real past customer message, then compare it with the one that was actually sent (four minutes).

If the helper's draft and the AI's draft both come out close to the original, the guide is doing its job for both. If a new person finds a rule confusing, a model probably does too, so rewrite it.

Where AI follows a voice guide too literally

  • Quirks become tics. A charity shop's guide said "light-hearted, a bit cheeky". Within a week every post ended on a pun, including one about a donated funeral outfit. The fix was the frequency limit from Step 3 plus an explicit "no jokes about bereavement, illness or money worries".
  • Banned words get synonyms. Ban "stunning" and you get "breathtaking". Ban the idea, not just the word: "don't praise appearance; describe the material instead".
  • Examples get copied. A subscription box company's guide included one great sample caption, and the model reused its opening ("Guess what landed in the studio this week?") in half the new posts. Include several examples per channel, and add "use examples for tone, never reuse their sentences".
  • Rules leak across channels. A casual Instagram rule turns up in a refund email. Channel notes, with the complaint row stated firmly, prevent most of this.
  • The voice hides missing facts. A well-voiced draft can still invent a lead time. Keep the "never invent prices, lead times or dates" rule in every version, including the compressed one.

Many of these are symptoms of the wider problem explained in why AI marketing copy sounds generic, and the prompt-side fixes are in writing AI prompts that sound like your business.

Keeping the guide alive

A voice guide that never changes drifts out of date as products, prices and customers change. Three habits keep it useful:

  1. Keep a change log at the bottom. One line per change: date, what changed, why. "1.2: added 'refined' to banned words after blind test."
  2. Add to the banned list whenever you edit a phrase out twice. If you keep deleting "we've got you covered" from drafts, the guide should say so.
  3. Review it each quarter against three recent pieces you're proud of and three AI drafts you had to rewrite heavily. The gap between them is next quarter's edit.
  4. List every place the guide is loaded at the top of the change log: for example the marketing Project in ChatGPT, a Claude Project used for product pages, and the Canva Brand Voice field. When the guide changes, update all of them the same day, or different tools will quietly write in different versions of your voice.

A final check that the guide is working: over a month, track how long AI drafts take to edit before publishing. For the furniture maker, an illustrative drop from about 12 minutes per caption to about 4 over two rounds of testing is the kind of change that shows the guide is doing its job. If editing time isn't falling, the guide is too vague, too long, or not loaded where people draft. For captions specifically, writing Instagram captions with AI that sound like you applies the same guide to one channel in detail.

Brand voice guide questions

How long should a brand voice guide for AI be?

One to two pages for the main guide, plus a separate file of example pieces. Longer guides tend to be followed less well, because rules buried on page five get less weight than the first few instructions, and contradictions creep in. If a tool limits the space, as Canva's Brand Voice does at 500 characters, keep a compressed version that holds the traits and the banned phrases.

Should the voice guide include examples of competitors' writing?

Only as clearly labelled "not us" examples, and sparingly. A short passage showing the tone you want to avoid can help the model draw a line. But competitor text in the examples file risks the model borrowing its phrases or structure. Your own best writing, plus your own before-and-after rewrites, teaches the voice far more reliably than someone else's.

Do I need a different voice guide for each AI tool?

No. Keep one master guide as a document you own, and paste or upload it into each tool: a ChatGPT or Claude Project, a Gemini Gem, the brand voice field in your design or email tool. Tools change and some are retired, as OpenAI's custom GPTs will be on 11 December 2026, so the guide should live somewhere you control, not only inside one product.

Further reads

Sources: OpenAI help pages on Projects in ChatGPT and the custom GPT retirement; Canva help pages on Brand Voice (checked September 2026).

Want AI drafts that actually sound like you?

On a 1:1 call we'll go through your best existing writing, turn it into a voice guide your AI tools can follow, and set it up where your team drafts so every tool uses the same one.

Book a 1:1 call with me