How to train an AI writing assistant on your brand voice

To train an AI writing assistant on your brand voice, give it a small, representative set of approved examples, organize those examples by voice and channel, and test the result on a real assignment. Do not begin by importing everything your company has ever published.

In kloudboard, this process uses the Brain, the workspace’s shared knowledge base. You can import social content, upload documents and media, add notes, and pin relevant folders as context for a specific reply. kloudie can then use imported videos to learn your speaking voice and imported social posts to learn your posting voice.

This is better understood as building a controlled context library than permanently retraining a foundation model. Your source selection still matters. If the library contains conflicting, outdated or generic material, the resulting draft can reproduce those problems.

Decide which voice you are teaching

“Brand voice” often combines several voices that should remain separate. A founder speaking on YouTube may sound conversational and opinionated, while the same company’s product announcements are concise and factual.

Start by choosing one of these targets:

• Speaking voice: scripts, podcast introductions, presentations and voiceovers.
• Posting voice: social posts, captions, threads and community updates.
• Editorial voice: articles, newsletters, reports and educational material.

You can teach all three, but do not mix them into one undifferentiated folder. A better structure is one folder per voice, with channel-specific subfolders where necessary.

If your immediate task is a 60-second founder-led video, begin with founder speaking examples. Adding formal reports will create more context, but not necessarily more useful context.

Build a representative source set

Open the Brain and assemble examples that reflect how the brand should sound now. kloudboard can import channels and posts from YouTube, TikTok, Instagram and X by link or connected account. Imported videos are transcribed, allowing kloudie to work from their spoken content.

You can also upload documents, spreadsheets, audio and video. Audio and video uploaded to the Brain are transcribed automatically. This is the right path for recorded voiceovers, internal presentations or unpublished interviews because audio and video files cannot be attached directly to a chat message.

A practical first set might contain:

• Three published videos that accurately represent the presenter’s current style.
• Ten approved social posts covering short, medium and longer formats.
• Two recent articles or newsletters.
• One page of explicit writing rules and prohibited claims.
• One note explaining the audience, offer and desired reader response.

These quantities are working assumptions, not measured requirements. The purpose is to provide enough variety to reveal recurring patterns without burying the strongest examples under marginal material.

Selection quality matters more than raw quantity. Exclude posts written for one-off campaigns, content produced before a major repositioning, and high-performing examples that do not sound like the brand you want to maintain. Performance can help you select material, but popularity alone does not make an example suitable for voice training.

Organize examples around decisions the assistant must make

A folder called “brand stuff” gives your team little guidance. Use folder names that tell a writer when the material applies.

For example:

• Brand voice
  • Founder speaking — approved
  • Social posting — approved
  • Editorial — approved
  • Claims and evidence rules
  • Archive — do not use

Inside each approved folder, add a short note describing the patterns the assistant should preserve. Do not rely on examples alone when a rule can be stated clearly.

A useful voice note could say:

Write in direct English. Open with the answer rather than a story. Use short paragraphs and concrete production examples. Explain tradeoffs. Avoid inflated adjectives, unsupported performance promises and rhetorical questions. Use technical language only when the audience needs it.

Also document channel differences. For example, an X post may open with a compact observation, while a YouTube script may need a clearer spoken setup. Asking every channel to imitate the same surface structure often makes the voice less natural, not more consistent.

Because the Brain is workspace-scoped, its contents are shared knowledge for that workspace’s members. Treat approved voice material as a maintained team resource rather than a private prompt saved by one writer.

Run a controlled voice test

Do not roll the setup out across your content calendar immediately. Pin the relevant Brain folder as context for one small assignment, then ask for a draft with a defined audience, channel, purpose and length.

Here is a reusable test request:

Use the pinned “Founder speaking — approved” folder. Write a 60-second video script for freelance video editors who lose time consolidating client feedback. The purpose is to explain one practical review rule. Keep the language conversational but not casual. Do not invent statistics, customer stories or product capabilities. End with the action the viewer should take on their next project.

Generate one draft first. Asking for many versions makes review slower and can hide whether the underlying voice match is sound. AI chat and generation use the workspace’s AI allowance, with costs varying by model, input, output and tools, so a focused test is also easier to scope.

Worked review example

Assume the draft opens with: “Transform your creative workflow with a revolutionary feedback system.” That sentence fails even if the grammar is correct. It uses broad promotional language, does not identify the problem precisely and conflicts with the instruction to start with a practical answer.

A better revision would be: “Set one deadline for consolidated feedback, and ask the client to appoint one person to send it.”

The revised version is more specific and actionable. It also gives reviewers something objective to assess: whether this direct opening resembles the approved examples.

Score the draft without reducing voice to personal taste

Have one editor and one person responsible for the brand review the same draft. Use a short scorecard rather than comments such as “make it punchier.”

Mark each criterion as pass, revise or fail. This is not a numerical brand-voice score. It is a way to locate the source of disagreement.

If tone passes but structure fails, improve the assignment instructions or add stronger examples from that channel. If unsupported claims appear, strengthen the claims note rather than merely deleting one sentence. If everything sounds generic, reduce the source set to its most distinctive examples.

Promote the workflow gradually

After the first test, run two more assignments in the same voice: one close to the examples and one on a new topic. This helps distinguish real voice consistency from imitation of a single source.

Then save the accepted instructions alongside the approved folder. A simple team handoff should identify:

• Which Brain folder to pin.
• Which content types the folder covers.
• Which claims require manual evidence checks.
• Who reviews the first draft.
• Where rejected or outdated examples belong.

Keep weak material out of the approved set. Move superseded examples into an archive clearly labeled not to use, or remove them when there is no reason to retain them. Review the set after a rebrand, a new presenter, a major audience change or a shift in editorial policy.

The expected outcome is not copy that can be published without judgment. It is a stronger first draft that starts from shared examples and explicit rules, giving writers and reviewers a common basis for revision. For a wider view of where the assistant fits in production—and where it does not—read the guide to kloudie for creative teams.

Start with one voice, one approved folder and one real deliverable. If the draft fails, diagnose whether the problem came from source selection, organization, assignment context or review rules. Fix that layer before inviting the whole team to use the system.