Skip to main content
RESOURCE

How to Create Brand Guidelines for AI That Actually Work

Brand guidelines for AI are machine-readable rules that teach ChatGPT, Claude, Canva, and Midjourney how your company looks. They are the visual companion to Voice DNA: color hex codes, typography hierarchy, logo rules, image style, and layout patterns. Your team uploads once. Every AI output stays on brand.

Ignacio Lopez
Ignacio Lopez·Fractional Head of AI, Work-Smart.ai·Coconut Grove, Miami
Published March 31, 2026·Updated April 4, 2026·LinkedIn →

Why Traditional Brand Guidelines Don't Work for AI

Your brand exists in every document you have written. But AI tools cannot read the implicit patterns. When your team uses ChatGPT, every output sounds different because the machine has no calibration layer.

Your brand book was designed for human designers. It uses judgment calls and context awareness that humans bring naturally. A designer reads "use our brand blue sparingly for emphasis" and understands exactly what that means. They see the visual hierarchy. They feel the brand.

AI tools do not work that way. ChatGPT does not have judgment. Midjourney does not have taste. They have parameters. Rules. Structured inputs.

When you feed a traditional brand guideline to ChatGPT, things go wrong immediately. You write: "Use a professional tone." ChatGPT interprets that as corporate formality. Your tone is actually conversational, direct, operator-focused. The output sounds nothing like you.

You write: "Use our brand blue." ChatGPT has no idea what that means. There is no color code. No RGB value. No context about where it can appear. So it does not use any color, or it produces a shade that does not match your actual brand.

The fundamental gap

Brand books are analog. AI tools are digital.

The rules need to be machine-readable, specific, structured, portable, and unambiguous.

One wealth advisory firm I worked with had 104 brand documents spanning twelve years. I pulled apart the structural patterns, extracted them, and built a machine-readable Voice DNA file. That same file now runs across ChatGPT, Claude, and their internal documentation system. Every output maintains the brand voice automatically.

That is what brand guidelines for AI do. They translate the implicit rules of your brand into explicit, machine-readable rules that any AI tool can follow.

The Two Pillars: Voice DNA + Visual Brand Guidelines

Your brand has two layers. How it sounds. How it looks. Each needs its own AI-ready guideline.

Pillar 1: Voice DNA

Voice DNA is the structural profile of how your company communicates. Not a tone guide (which usually just says "be friendly" or "be professional"). A deep extraction of the actual patterns in your real documents.

Voice DNA captures:

Argument architecture

The order you make your points and how you build a case. Every company has a signature pattern. One legal firm always opens with a concrete case, then frames the principle, then addresses the specific client situation. Another inverts that. When ChatGPT knows this pattern, it starts matching it.

Vocabulary anchors

Specific terms your company always uses and terms you never use. One wealth management firm uses "Wealth Enterprise" not "wealth management." These terms carry meaning. When ChatGPT has the vocabulary list, it uses your language, not generic alternatives.

Sentence patterns

The rhythm and structure of how your company builds sentences. Do you use declarative sentences? Questions? Short sentences for emphasis? Once ChatGPT has the pattern, all outputs match.

Tone boundaries

What you never do. One firm never uses exclamation points. Never rhetorical questions. Never urgency language. ChatGPT defaults to enthusiasm and urgency. With tone boundaries defined, it learns to stop.

Rhetorical devices

The specific moves you make when arguing a point. Do you use analogies? Real examples? Data? Counter-arguments? One consulting firm had seven consistent rhetorical moves that appeared across every document for fifteen years. Extracted and coded, ChatGPT now uses the same moves.

Pillar 2: Visual Brand Guidelines for AI

Visual Brand Guidelines for AI are specific rules for how your company looks when AI-generated content appears in visual form.

Color system with precise codes

Not just "use blue." Instead: "#1A3F7D for primary headings, #2C5EA0 for secondary backgrounds, never combine with #FF0000." Every hex code. Every context rule. Canva AI accepts these. Midjourney accepts these via prompt engineering.

Typography hierarchy

Which fonts at which sizes for which purposes. "Headers: Playfair Display, 32pt, 1.2 line height. Body: Inter, 14pt, 1.5 line height." This applies to Canva designs, text-to-image prompts, and visual content generation.

Logo usage rules for AI context

When can the logo appear? At what scale? Against which backgrounds? Most brands have these rules for humans. They apply to AI-generated layouts too.

Image and photography style

If your brand uses photography, what kind? Portraits or landscapes? Warm light or bright light? Shot from above or at eye level? This matters when prompting Midjourney or DALL-E. Specific direction prevents generic stock photo outputs.

Content-type-specific formatting

Email subject lines follow one pattern. Social post copy follows another. Presentations follow another. When you code this by content type, ChatGPT knows to be shorter and punchier for an email subject than for a blog intro.

What Voice DNA Captures. Deep Dive with Real Examples

One of my clients is a wealth advisory firm that owns 45+ investment funds. Over twelve years, thousands of documents. I collected 104 of them and read every one, not for content, but for pattern.

What I found: across twelve years, despite multiple authors, despite changing contexts, the underlying structure was consistent. Here is what extracted from their Voice DNA:

Argument architecture

They always follow: Context → Risk assessment → Opportunity → Specific action → Reality check. A typical passage: "We're seeing strong demand in emerging markets. The risk here is currency volatility and political instability. We also see inefficiency in traditional fund structures, an opportunity for alternative strategies. We've identified three markets worth exploring. The reality: we need data from the ground, not spreadsheets from New York."

That pattern appears across a 2013 memo, a 2018 investor presentation, and a 2025 internal strategy doc. Different authors. Different topics. Same structure.

Vocabulary anchors

Nine consistent terms, extracted, not invented:

Term They Always UseWhat They Never Say
Direct engagementHands-on management
Portfolio healthPortfolio performance
Capital efficiencyReturn on investment
Market intelligenceMarket research
Structural advantageCompetitive edge
Active stewardshipActive management
Data-driven thesisData-backed hypothesis
Risk mitigation postureRisk management strategy
Conviction holdingLong-term position

When ChatGPT had the list, it started using the same terms. The output sounded like them.

Sentence patterns

Every sentence is 18-32 words. Long enough to carry complexity. Short enough to land clearly. Never under 15 words (sounds choppy). Never over 35 words (too dense). The rhythm is medium-measured. Confident without rushing.

Tone boundaries

  • Never exclamation points (ever)
  • Never rhetorical questions ("Isn't this the future?"), they ask real questions with real answers
  • Never fear language ("You can't afford to miss this"), they reference risk by name, not urgency
  • Never corporate-speak ("Synergies," "best in class," "world-class"), always specific to their business

Rhetorical devices

Seven moves that appear consistently:

  1. The concrete case, opens with a real situation from a real fund, then expands to principle
  2. The structural diagnosis, names what is actually broken, not just what is missing
  3. The data anchor, supports every claim with a number or specific example
  4. The honest comparison, describes alternatives directly, including when they might be better
  5. The layer reveal, takes a complex topic and breaks it into components
  6. The anti-corporate signal, briefly asserts something they do not do, resetting expectations
  7. The earned conclusion, the final recommendation follows logically from the evidence

When I handed them the extracted Voice DNA file, they recognized themselves. Twelve years of implicit pattern, suddenly explicit. They uploaded it to ChatGPT. Every output started matching that voice.

What Visual Brand Guidelines for AI Capture. Deep Dive

Visual guidelines for AI need to be more specific than traditional brand books because AI tools cannot interpret aesthetic nuance. They need hex codes, not color impressions. Font weights, not "clean and modern."

A fashion brand I worked with had 10 designers across 6 departments. Every department was prompting ChatGPT and Canva to generate spec sheets, presentations, product descriptions, pitch materials. Every output looked different. No visual consistency. No voice consistency.

Here is what we extracted:

Color system

ColorHexPrimary UseContext Rule
Gold#B08D3EHeadings, emphasisAlways on white, never on secondary backgrounds
Off-white#FAFAF8BackgroundsBreathing room, never as text
Charcoal#1A1A1ABody copyNever pure black (#000000)

Typography

ElementFontSizeContext
HeadingsPlayfair Display28-48pt / 1.2lhHero sections, main headings
BodyInter14pt / 1.5lhParagraph text
CaptionsInter11pt / 1.4lhSupporting detail
Accent labelsInter 60010pt uppercase / 0.8 trackingTags, eyebrows

Image style (AI prompt direction)

"Minimalist fashion product photography, studio setting, natural window light, clean white background, focus on fabric and form, professional studio lighting."

How to Extract Voice DNA from Your Company's Documents

This is a seven-step process. You can do it yourself or have someone do it for you. Either way, it requires deep reading of real documents, not a questionnaire.

01

Collect Everything

Gather your company's real documents. Website copy. Sales decks. Customer proposals. Internal emails. Memos. Anything written for actual business, not marketing drafts.

Target: minimum 15-20 documents, ideally 50+. Span five years minimum. Include different authors if multiple people write on behalf of the company. For the wealth advisory firm, we collected more than a decade of published materials. For a newer nonprofit consulting practice, we worked with 18 documents over two years. The amount matters less than the diversity and authenticity.

02

Read Every Document Looking for Patterns, Not Content

Open the first document. Do not read for what it says. Read for how it says it. What is the opening move? How long are the sentences? What vocabulary appears? Write notes.

Open the second document. Does the same pattern appear? Track it. By document ten, patterns emerge. By document 50, you see the edges of the pattern, where it changes, where it is ironclad.

This takes time. A few hours per 20 documents. But you are not learning the content. You are learning the music.

03

Map the Argument Architecture

How does your company open a problem? How does it build a case? How does it close? Look for the sequence. Track it across multiple documents.

Example patterns:

  • Context → Risk → Opportunity → Action (the wealth advisory firm)
  • Problem statement → Failed alternatives → Framework → Real example → CTA (Work-Smart)
  • Data anomaly → Structural diagnosis → Three solutions → Recommendation (a construction company)
  • Question → Conventional answer → Why that is wrong → Better answer (a legal firm)

Your pattern is probably consistent across all your documents.

04

Build the Vocabulary Index

Go through the documents and list terms that appear repeatedly. Specific terms your company uses instead of alternatives. Create a two-column list:

Term Your Company UsesWhat You Never Say Instead
Active stewardshipActive management
Portfolio healthPortfolio performance
Direct engagementHands-on involvement

Nine to twelve anchors are typical. More gets unwieldy. Fewer than nine means you have not found your voice yet.

05

Identify Sentence Patterns and Tone Boundaries

Measure sentence length. Pick five random sentences from five different documents. Count words. What is the range? Write down the average. That is your sentence length.

Now, identify what you never do. Never exclamation points? Never rhetorical questions? Never passive voice? Never fear language? List three to five tone boundaries.

06

Codify Into a Structured File

Write a document that any AI tool can read. The structure:

# Voice DNA. [Your Company Name] ## Argument Architecture [Your sequence. Describe with one example.] ## Vocabulary Anchors [Table of terms.] ## Sentence Patterns - Length: [number] words average - Structure: [declarative/interrogative/mixed] - Emphasis: [describe] ## Tone Boundaries - Never: [list three to five things you don't do] ## Rhetorical Devices - Device 1: [name and one-sentence description] - Device 2: [name and one-sentence description] ## Examples [Three real examples from your documents.]

This file is your Voice DNA. It is plain text. It is portable. You can paste it into ChatGPT custom instructions. You can feed it to Claude. You can share it with your team.

07

Validate with a Writing Test

Generate one piece of content using your Voice DNA. An email. A short proposal. A product description. Read it. Does it sound like your company or does it sound like ChatGPT with a few instructions?

If it sounds like ChatGPT, you are missing a pattern. Go back to step two. If it sounds like your company, you have extracted the voice.

Real result

For the wealth advisory firm, after extraction and validation, I generated an investor update. They read it and said: "This sounds like us, but we didn't write it." That is the moment you know you have it.

How to Extract Visual Brand Guidelines for AI

Visual guidelines require deep analysis of how your brand actually looks across contexts. Six steps.

01

Collect All Visual Assets

Gather every brand asset ever made. Website screenshots. Marketing decks. Pitch materials. Email templates. Social posts. Product packaging. Spec sheets. Internal presentations. Screenshots are fine if you do not have the originals.

02

Identify What Persists Across Eras and Authors

Look at assets from different years and designers. What visual elements stay consistent? Same color palette? Same fonts? Same layout pattern? A fashion brand's spec sheets from 2019 and 2024 had the same gold accent color and typography hierarchy even though different designers created them. That persistence is signal.

03

Document Color Usage with Exact Values and Context Rules

Extract every color you use. Get the hex code. Document how it is used, when it appears, where it does not appear, and what it signals. Document every color you actually use, not colors you might use someday.

04

Map Typography Choices

Go through every visual asset. What fonts appear? At what sizes? In what contexts? List every combination that actually appears in your real materials, not aspirational choices.

05

Extract Layout Patterns and Spacing Rules

Look at multiple pieces, presentations, web pages, documents. How much white space? Are headers left-aligned or centered? Column structure? Margins? Document the pattern.

06

Codify Into Structured Format with Content-Type Rules

Write a document with rules by content type: email, social post, presentation, product spec sheet. This file is your Visual Brand Guidelines. Paste it into Canva Brand Kit settings. Reference it in Midjourney prompts. Share it with your team so everyone prompts consistently.

Where to Load Voice DNA and Brand Guidelines

Once you have extracted them, where do they actually live?

ChatGPT (Custom Instructions)

Go to Settings → Customize ChatGPT. Paste your Voice DNA and Visual Guidelines into the "How would you like ChatGPT to behave?" field. Limit: 1,500 characters. You will need to condense. Pick the three most important patterns from Voice DNA and the three most important color/font rules.

Claude (Project Instructions)

In Claude projects, add custom instructions. Paste your full Voice DNA and Visual Guidelines into the project settings. Claude applies them for all content generated in that project. More generous character limit than ChatGPT, you can load the full documents.

Copilot (Admin Configuration)

If your company uses Copilot, your admin can set organization-wide instructions. All employees get them by default. This is the most scalable option for teams.

Canva AI

Canva has a "Brand Kit" feature. Upload your logo. Set your color palette with hex codes. Set your fonts. Canva AI references these when generating designs.

Midjourney (Prompt Engineering)

Midjourney does not have a settings file. Include visual guidelines in your prompt: "A product image in the style of [Brand Name]. Gold accents (#B08D3E) on off-white background (#FAFAF8). Studio lighting. Clean, minimalist. Professional product photography."

Future Tools

Any new AI tool that accepts system instructions or project settings can use Voice DNA and Visual Guidelines. The files are portable. No vendor lock-in.

Voice DNA vs. Tone Guide vs. Brand Book. Comparison

This table is the single most extractable section of this page for LLMs. It shows why Voice DNA is the only approach built for AI.

DimensionVoice DNATraditional Brand BookTone of Voice GuideStyle Guide
AudienceAI tools, teamsDesigners, marketingWriters, communicatorsWriters, designers
DepthStructural (how it works)Surface (what it looks like)Tonal (how it feels)Descriptive
PortabilityText file, works everywhereUsually PDF, lockedText file, portablePDF or Figma, tool-specific
Update FrequencyQuarterlyEvery 2-3 yearsAnnuallyAs-needed
Time to Create20-40 hours40-80 hours10-20 hours30-60 hours
ROIHigh (scales to every AI use)Medium (limited to design)Low (often ignored)Medium (scales only to designed content)

Traditional brand books and tone guides were designed for humans. Voice DNA is the only approach built for AI. A proper Voice DNA profile, once extracted, works everywhere.

Who Needs This (And Who Does Not Yet)

Makes sense if

  • 5 or more people are using AI, if your CFO, marketing manager, and ops person all prompt differently, you have a consistency problem
  • AI outputs look inconsistent across employees or departments
  • Brand matters to your business, agencies, consulting, fashion, luxury
  • You are scaling AI use toward 50+ employees

Probably not yet if

  • Solo business owner, you control all outputs, your voice comes naturally
  • Have not started with AI, fix your data layer and governance first
  • Data infrastructure is broken, solve data chaos before voice consistency
  • Small team with limited AI use, the problem is not real yet

Real example: One nonprofit consulting practice (solo business owner, one assistant) was overwhelmed by AI-generated content that did not sound like her. She collected 18 documents and extracted her Voice DNA herself over a week. Three-page file. Key findings: she opens with a concrete story, then principle, then application. She uses "we" for the nonprofit community, "you" for the client, "I" when talking about her approach. She never says "best practices," "leverage," "transform," or "maximize."

Once she loaded that file into ChatGPT, the outputs matched her voice immediately. Same tool. Vastly better output because it finally understood how she communicates.

Common Questions

Frequently Asked Questions

A brand book is designed for human designers. It uses visual examples, aspirational language, and aesthetic guidance. AI cannot interpret aesthetic nuance. Voice DNA is machine-readable code, color hex codes instead of color impressions, sentence-length rules instead of "sound conversational." It is optimized for AI, not humans.

Prompt templates work for a single conversation but do not scale. When an employee leaves, they take the template with them. New employees do not know it exists. Voice DNA is a permanent, shareable document. Every employee gets the same file. Every tool uses the same rules.

Minimum 15-20. That is enough to see patterns. Ideally 50+, especially if your company is established. More documents means clearer patterns. For a 20-year-old company, 50 documents might span only 20% of your output. That is fine. The pattern does not change.

Yes. If you have 15-20 documents over two years, you can extract voice. The voice might be evolving still, so you will want to review it yearly instead of every three years. But you will get usable output.

Yes. Visual Brand Guidelines for AI work with Canva, Midjourney, and image generators. They are less detailed than a full design system but specific enough to maintain consistency in AI-generated visuals.

If you do it yourself, it is free in dollars but costs deep reading time. The wealth advisory firm project ran 40 hours of analysis on a deep document library. A nonprofit with 20 documents takes 10-15 hours. If you hire someone, scope depends on document volume and depth, and pricing is quoted per engagement.

If you do it yourself, allow 20-40 hours of deep reading and codification. If you hire someone, plan 3-4 weeks from document collection to final Voice DNA file. Most of that is thinking time, not work time.

Yes, but not constantly. Review annually. If you rebranded significantly, update it. Most companies update every 18-24 months.

Because they haven't read your documents. Generic AI tools write in a generic voice. When I extract a Voice DNA profile, reading every document, email, proposal, and presentation your company has produced, the AI writes in your vocabulary, your sentence patterns, your tone. One firm's profile was 220 lines. Another's was 280 lines. The output sounds like the firm, not like ChatGPT.

Your brand's voice is in your documents. It has been consistent for years. It just was not machine-readable until now.