How to Brief AI on Your Brand Voice

You've figured out that your team's AI outputs all sound different. One person's ChatGPT email sounds like a consultant wrote it. Another person's Claude draft sounds like a LinkedIn guru. A third person's Copilot output is technically correct but reads like a government form. You know this is a voice problem. You know the fix is to brief the AI properly. What you don't know is how to write that brief.

This is the gap I see most often. Companies have a brand voice document somewhere (maybe a PDF from an agency, maybe a Notion page, maybe just in the founder's head). The document wasn't written for AI. It was written for human creative directors. When employees copy-paste it into ChatGPT, the AI can't actually use it, because it's written in the wrong format.

The fix is a brand voice brief structured specifically for AI consumption. I've written these for construction firms, fashion brands, wealth advisors, and nonprofits. Same six-section structure every time. Here's the anatomy.

Section 1: Identity and audience. Two paragraphs max. Who the company is in one paragraph, who the audience is in another. The identity paragraph names the category (construction, wealth advisory, wellness brand), the positioning (premium, operator-led, accessible), and the founder voice if relevant. The audience paragraph names who the content is written for: their role, their context, their stage of buying, and what they care about.

Bad example: "We are a dynamic, innovative firm committed to excellence." That's meaningless to an AI and useless to a human.

Good example: "Work-Smart is a solo consulting practice offering fractional AI leadership to mid-market operators ($5M-$100M revenue). Founder Ignacio Lopez speaks directly to CEOs, COOs, and Managing Partners. Audience is operators who have been burned by a prior tech project and are skeptical of consultants. They respond to specific examples, honest pricing, and the word 'operator' used as a term of respect."

Section 2: Structural voice rules. The non-negotiables. Sentence structure, paragraph length, use of formatting. This is where most brand guidelines fail, because they use vague words like "confident" or "authoritative" or "approachable." An AI can't act on those. An AI can act on: "Sentences average 14 words. Never use em-dashes or en-dashes. Use colons, commas, or periods for pauses. Paragraphs are 2-4 sentences. No exclamation points. No rhetorical questions as openers."

The test is: can a non-native speaker with no context apply these rules? If yes, you've written them correctly. If the rules require interpretation, rewrite them.

Section 3: Vocabulary. Two lists. Allow list (words and phrases your brand uses deliberately). Ban list (words and phrases your brand never uses, with brief reasoning for each ban).

Allow list examples from real client briefs: "operator" (as a term of respect for business owners). "Ship" (as a verb meaning deliver). "Production-ready." "Fractional." "Mid-market." "Retainer."

Ban list examples from real client briefs: "Transform / transformation" (banned because overused). "Leverage" (corporate speak). "Game-changing" (marketing fluff). "Best-in-class" (meaningless). "Synergy." "Thought leadership." "In today's fast-paced world." Any word ending in "-ify" (gamify, verify, solidify, unless it's a common word).

You don't need 200 words on each list. You need the 15-25 that matter most for your category.

Section 4: Sentence-level patterns. This is the section that separates a good brief from a generic one. Describe the rhythm, not just the rules. Give the AI the patterns your best writers use without thinking.

Examples from client briefs:

"Use the condition, reality, what-to-do pattern. State the condition the reader is in. State the reality. State what to do. Example: 'Your team is using ChatGPT without oversight. Most of them are on the free tier. Publish a one-page policy this week.'"

"Open with a specific, not an abstraction. Not 'AI is changing the workplace.' Instead: 'Your three-person legal team is drafting NDAs manually.'"

"Use the two-beat structure for complex ideas. First beat states the common view. Second beat states your position. Example: 'Most consultants will tell you to start with a strategy deck. Don't. Start with the data.'"

Every good brand has 3-5 patterns like these. Your job is to name them explicitly.

Section 5: Approved and rejected output examples. Two or three paired examples. For each: an approved output and a rejected output on the same prompt. Explain in one line why the approved version works and why the rejected version fails.

Example:

Prompt: "Write a one-line email subject about a new AI feature."

Rejected: "Transform Your Workflow with Our Revolutionary New Feature." (Why it fails: uses "transform," "revolutionary," overpromises. Generic.)

Approved: "The new calendar view now runs on the dashboard." (Why it works: specific, declarative, no hype, tells the reader what changed.)

The AI learns patterns from examples faster than from rules. Four to six paired examples in this section usually raises output quality more than any amount of additional rules.

Section 6: Failure modes and recovery. What the AI does when it drifts. Most AI outputs drift toward a default style: overly formal, over-adjectived, full of transitions, heavy on corporate speak. Name the drift. Give the AI a specific instruction for what to do when it catches itself drifting.

Example: "If a draft contains any of the banned vocabulary, any em-dashes, or any exclamation points, revise before delivering. If a draft runs more than 20 words per sentence on average, break it into shorter sentences. If a draft opens with 'In today's,' rewrite the opening."

This section is the safety net. It catches the drift before the output reaches a human editor.

A brief is only useful if it's loaded where the AI will see it. Three common locations:

ChatGPT Custom Instructions. Paste the brief into the custom instructions field, which gets loaded automatically with every new conversation. Best for individual use. Every team member does this once and all their ChatGPT outputs improve.

Claude Projects or Custom Instructions. Claude Projects let you load the brief as project knowledge, which stays active across all conversations in that project. Best for team use, because you can share a project with everyone.

Microsoft Copilot Agents. Copilot supports custom agents where you can paste the brief into the instructions. Particularly useful inside Microsoft 365, where the agent shows up inside Outlook, Word, and Teams.

CLAUDE.md or system prompt file. For developer workflows and Claude Code, the brief lives in a CLAUDE.md file at the project root. Every new session loads it automatically. This is how Work-Smart keeps voice consistent across every client repository.

One brief. Four deployment locations. Same output quality regardless of which tool an employee picks up.

Target 800-1,500 words. Shorter is not worse, if every word earns its place. Longer almost always underperforms, because the AI starts treating the brief as background context rather than active rules.

If your current brand document is 12 pages of PDF, that's normal. It was written for humans. The AI version is the tight distillation. Keep the PDF for onboarding new hires. Keep the 1,000-word version for the AI.

Three triggers for an update.

You add a new service or audience. The identity and audience section needs to be amended, because the brief should reflect the current company.

You catch a systematic drift. If three of your team's AI outputs in the same week share the same off-brand pattern (all overly formal, all using a specific banned word), that's a signal the brief is missing a rule. Add the rule. Re-load the brief.

You rebrand. Obvious, but often missed. When the brand shifts, the brief shifts. The voice stays consistent with the brand, not the other way around.

Five patterns I see consistently at mid-market companies attempting this for the first time.

Mistake one: too many adjectives, not enough rules. "We're confident, warm, expert, accessible, and authentic." That's a personality description, not a brief. The AI can't act on it. Replace every adjective with a specific rule.

Mistake two: banning without examples. "Don't use corporate jargon" is not a useful rule. "Don't use: leverage, synergy, best-in-class, touch base, circle back, thought leader" is useful. Name the specifics.

Mistake three: showing only approved examples. Without rejected examples, the AI doesn't learn the boundary. It learns the target but not the edges. Always include rejections, and explain why they fail.

Mistake four: writing the brief in the voice of the brief. If the brand is operator-led and direct, the brief should be operator-led and direct. If the brand is warm and accessible, the brief should be warm and accessible. The brief is itself an example of the voice.

Mistake five: treating it as a one-time deliverable. The brief improves every time you use it. Keep a note of outputs that surprised you (good or bad) and update the brief accordingly. After 3-6 months, it's far better than version one.

Brand voice is the single biggest AI consistency lever, and it's usually the most ignored. Every team that gets this right stops arguing about AI outputs in email threads. Every team that gets it wrong keeps losing hours editing the same drift patterns over and over.

If you want to see what a voice brief looks like for a company at your size, the Voice DNA extraction service is the fastest path. Two to three weeks. Fixed fee. You walk away with a brief loaded into ChatGPT, Claude, or Copilot, plus the Brand Guidelines for AI companion document (visual identity rules formatted so AI tools can apply them: colors, typography, logo use, image style, layout patterns). For the symptom you are probably seeing right now, read AI Brand Consistency for Team Outputs, and for the structural companion piece, how to build an AI style guide your team actually uses. Or take the free assessment to see where you stand first.