VibeContent

Try “Veo”, “voiceover”, “thumbnail” or “Descript” · Esc to close

🏢 Teams & agencies · Guide 2 of 2

Keeping One Brand Voice Across Every AI Tool

9 min read · Last reviewed 25 Sep 2026

A team of ten may write with Jasper, draft in Claude or ChatGPT, lay out in Canva and caption in a video editor. Freelancers bring their own tools on top. Each tool has its own idea of your brand, or none at all. The result is copy that is correct and on-topic but sounds like four different companies. A client can spot that faster than any checker.

This guide covers how to keep one voice across all of that. It explains where the voice should live, how to carry it into each kind of tool, how to review and measure it, and what to do when a model update quietly changes the output.

Keep one source of truth, outside every tool

Most brand-voice problems start with the voice being defined in five places. There is a Brand Voice in one app, a pasted paragraph in someone's custom instructions, a PDF from the agency that did the rebrand, and a few habits in the head of the senior writer. They drift apart within a month.

Write the voice once, as a plain document the team owns, stored in version control or a shared drive with a change log. Every tool setting is then a copy of that document, and you update the copies when the source changes. The source document should hold:

  • Three to five voice traits, each with a "this, not that" pair. "Direct, not blunt." "Warm, not cute." Traits without counter-examples are too vague to use.
  • Vocabulary rules. Words you always use (product names, capitalisation), words you never use, and terms reserved for legal or regulated claims.
  • Mechanics. Sentence length, British or American spelling, Oxford comma, emoji policy, how you write numbers and dates.
  • Channel adjustments. How the voice shifts between a LinkedIn post, a help article and a 15-second short video script.
  • Six to ten real examples. Approved pieces, each with a line on why it works, plus two or three rejected drafts with the reason they were rejected.

The examples matter more than the adjectives. Models copy examples well and interpret adjectives loosely.

Brand kits and brand-voice features: what they cover

Many apps now ship a brand feature, but they store different things. In the directory's listings:

Tool typeExamplesWhat the brand feature holds
AI writing platformsJasper, Copy.aiStored brand voices, product knowledge and audience profiles that shape generated copy
Design toolsCanva, Picsart, Kittl, LookaLogos, colours, fonts and templates (visual identity)
Video and ad toolsPictory, Creatify, Pencil, VizardBrand kits applied to video templates and ad variations
Avatar videoHeyGen, Synthesia, ColossyanBrand kits for on-screen styling

Two things vendors don't make obvious. First, most "brand kits" are visual: colours and fonts, not tone. Only the writing platforms really store voice. Second, voice slots are metered. At the time of writing, Jasper Pro includes 2 Brand Voices, 5 Knowledge assets and 3 Audiences per plan, and unlimited brand assets need the custom-priced Business plan. An agency with twelve clients hits that limit fast. So plan per-client voices before you buy seats.

When you fill a vendor's brand-voice form, paste from the source document. Don't let the tool "learn" your voice from a website scrape and then edit the result inside the app. That creates a second source of truth that nobody reviews.

Carry the voice into assistants with Agent Skills

General assistants are where voice drifts most, because everyone prompts them differently. Agent Skills are a reliable fix. A skill is a folder with a SKILL.md file: a short YAML header (name and description) followed by instructions, plus any reference files you add. The assistant loads only the name and description at the start. It reads the full instructions when a task matches, and it opens bundled files only when it needs them. So a detailed style guide costs almost nothing until it is used.

The format began at Anthropic and is now an open standard. The agentskills.io site lists support in Claude, Claude Code, ChatGPT and Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code and many other agents. One skill folder can serve most of your team's assistants.

A brand-voice skill that works looks like this:

brand-voice-acme/
├── SKILL.md            # traits, rules, review loop (keep it short)
├── reference/
│   ├── vocabulary.md   # always/never word list
│   ├── channels.md     # LinkedIn, email, help centre, video scripts
│   └── examples.md     # approved and rejected pieces with reasons

Anthropic's authoring guidance has advice that suits style guides well:

  • The description decides when the skill fires. Say what it does and when to use it, in the third person. For example: "Applies Acme's brand voice and vocabulary to marketing copy. Use when drafting or editing any Acme blog post, email, ad or social post."
  • Keep the SKILL.md body under 500 lines, and keep reference files one level deep so they are read in full.
  • Include a feedback loop: draft, check against the checklist, revise, and only then return the copy.
  • Build at least three evaluations before you write a lot of instructions, and test with every model you plan to use.

The limits matter for teams. Anthropic's docs say custom skills don't sync across surfaces. On claude.ai, custom skills are individual: each person uploads them, and admins can't manage them centrally. On the Claude API they are shared workspace-wide. In Claude Code they live in a personal or project folder, or ship as a plugin. Store the skill in one repository and treat each upload as a deployment, with a version number in the file.

You don't have to start from scratch. Anthropic's public skills repository includes a brand-guidelines example. The Marketing Skills pack uses a pattern worth copying: a shared product-marketing context skill that every other skill reads first, so copy, email and ad skills all start from the same positioning.

Only install skills from sources you trust, and read them first. A skill can tell an agent to run code. For design, the Canva MCP server lets Claude, ChatGPT and other assistants create designs in your Canva account. Brand kits and brand templates in that connector need a paid Canva plan.

Prompt libraries for everything else

Not every tool accepts skills, and not every task is worth one. For the rest, keep a shared prompt library, with each prompt starting from the same voice block:

  • Voice block: a condensed 150-250 word version of the source document, pasted at the top of any prompt in tools without brand features.
  • Task prompts: one per recurring job (newsletter intro, product launch post, case study outline), each with an approved example.
  • Owner and version on every prompt.

Public libraries are good raw material. prompts.chat is CC0 and can be self-hosted as a private team library. Fabric stores its patterns as plain Markdown files you can fork into your house style. Treat public prompts as templates. Many are generic or were written for older models.

Review and approval workflow

Tools lower the cost of a first draft. They do nothing for the cost of a bad publish. A workflow that scales to 5-50 people:

  1. Brief names the channel, audience and the voice version in use.
  2. Draft is made in an approved tool, with the brand skill, brand voice or voice block loaded.
  3. Self-check: the writer runs the draft through the same checklist the skill uses.
  4. Voice review by a named editor per client or brand. Not "whoever is free".
  5. Compliance review for claims, testimonials and AI labels (see our team policy on rights and disclosure).
  6. Approval logged with the voice version, so drift can be traced later.

Keep AI out of step 4 as the final judge. A model is useful for flagging problems, like banned words, sentence length or spelling variant. It is poor at deciding whether something sounds like you, because it tends to rate its own style as good.

Measure drift instead of arguing about it

"It doesn't sound like us" is a feeling. Turn it into numbers you can track every month:

  • Rule violations per 1,000 words: banned terms, wrong spelling variant, missing product-name capitalisation. A script or a simple checker can count these.
  • Average sentence length and reading level, compared with your approved examples.
  • Edit distance at review: how much the editor changed. If it rises, the drafts are drifting.
  • Blind test each quarter: mix five approved human pieces with five AI-assisted ones and ask two people outside the team to rate them against the voice traits.
  • A fixed test set: ten standard briefs you re-run in every tool after any change, with outputs saved side by side.

The test set is the most useful of these, because it is the only one that catches changes you didn't make.

Onboarding freelancers

Freelancers are where voice consistency usually breaks. On day one, give each one:

  • The source document and the current skill folder or voice block, with a version number.
  • A list of approved tools and the account they should use (a company seat, not their own, for the rights reasons in our policy guide).
  • Three briefs from the test set as a paid trial, reviewed against the checklist.
  • A clear rule: no pasting client material into tools that aren't on the approved list.

If freelancers use their own assistants, send the skill folder rather than a PDF. Because the format is an open standard, it works in most agents they will already have.

What breaks when a model updates

Model updates are the silent cause of drift. Vendors change the default model behind an app, and the same prompt now produces longer sentences, more hedging, or a new favourite word. Things that commonly break:

  • Length and structure. A new model may add headings, lists or summaries your voice doesn't use.
  • Instruction weight. Rules that were enough for one model get ignored by another, or followed too literally. Anthropic's skill guidance says to test with every model you plan to use, because what works for a larger model may need more detail for a smaller one.
  • Multi-model apps. Apps that route to several models can change what you get when someone picks a different model from a menu. Check the model map to see which models sit behind each app, and pin the model in team settings where the app allows it.
  • Vendor brand features. A brand voice trained on one model version may be applied by a newer one without warning.

The defence is routine. Watch the changelogs of your core tools. Re-run the fixed test set after any announced update, and at least monthly. Keep the previous outputs, so the comparison is easy. Adjust the skill or voice block, bump the version, and redeploy it to every surface.

Takeaways

  • Define the voice once, in a document you own. Every tool setting is a copy.
  • Most brand kits are visual. Only a few tools store tone, and they meter it per plan.
  • Agent Skills carry the same style guide into most assistants, but custom skills don't sync across surfaces, so deploy them on purpose.
  • Keep a human editor as the final voice check and log the voice version on each approval.
  • Measure drift with a fixed test set, and re-run it whenever a model changes.

For automation patterns, see the developer guides.

Read the official docs for…

Mentioned in this guide

The rest of this level

  1. Rights, Watermarks and AI Disclosure: A Team Policy
  2. Keeping One Brand Voice Across Every AI Tool

All four levels →