Brand constitution vs. brand guide
Most brand guides are written for humans. They describe tone with adjectives, show example headlines, and explain the brand's personality in paragraphs designed to be absorbed and internalized over time. A person reads them and develops judgment. That judgment produces on-brand work.
AI does not develop judgment. It follows patterns. Give it a 40-page brand guide and it will produce content that sounds adjacent to your brand — competent, plausible, and subtly wrong in ways that are hard to articulate and harder to fix at scale.
The document structure is the problem, not the AI.
What a brand guide is designed to do
Brand guides are built for human onboarding. They explain the brand through description: "We are warm but not casual. Confident but not arrogant. Direct but never blunt." The reader — a designer, a copywriter, a new hire — absorbs those descriptions and builds a mental model. When they write, they apply that model with implicit judgment about context, register, and stakes.
This works because humans fill gaps automatically. When a brand guide says "avoid corporate jargon," a good writer knows what that means in the context of a benefits FAQ versus a product launch announcement. They adjust.
AI has no gap-filling that serves brand intent. It has a distribution of outputs shaped by training, and it will produce content that satisfies the most plausible interpretation of your prompt — not the most brand-aligned one. Descriptive guidance does not constrain that distribution. Directive rules do.
What a brand constitution is designed to do
A brand constitution is a compact set of named, specific rules. Not descriptions of how the brand feels — directives specifying what the brand does and does not do. Each rule is named so it can be referenced. Each rule is specific enough that its application is unambiguous.
The structural difference looks like this:
| Brand guide | Brand constitution |
|---|---|
| "We use plain language" | Plain-language rule: No noun strings longer than 3 words. |
| "We're direct" | Directness rule: Lead with the conclusion. Never bury the claim in a subordinate clause. |
| "We avoid jargon" | Terminology ban: Do not use "leverage," "ecosystem," or "synergy." |
| "Our tone is confident" | Confidence rule: State claims as assertions, not hedges. Replace "may help" with "reduces." |
The left column describes the target. The right column specifies the behavior. A human can work from description; AI cannot. And critically, a human operating at scale — producing twenty pieces a week with AI assistance — cannot reliably translate description into consistent application either. The constitutional model is not just an AI fix; it is a scaling fix.
Why named rules work better than descriptive guidance
Naming rules has two practical effects. First, it makes them enumerable. You can list a brand's core rules in a handful of lines — small enough to include in every content generation prompt, specific enough to actually constrain output. Second, names make violations detectable. "The output violates the plain-language rule" is a specific editorial note. "This doesn't feel on-brand" is not.
Research on instruction-following in large language models establishes why this matters technically. A compact set of named principles enables consistent adherence in a way that descriptive guidance does not — the model can check each rule explicitly against its output rather than attempting to satisfy a diffuse tonal description. The same insight behind Constitutional AI applies here: compact named principles, not exhaustive description, is the architecture that produces reliable behavior.
For marketing teams, this is the finding worth acting on. The constitutional model is not a workaround for AI limitations — it is a better way to encode brand intent for any output channel, human or AI.
The compact that replaces the guide
A brand constitution does not need to be long. It needs to be complete within a narrow scope: the rules that actually determine whether AI output sounds like you. That is rarely more than a page.
A working constitution typically has five components:
- Terminology bans — words and phrases that are off-brand, regardless of context
- Required terms — words and phrases that should appear, and where
- Voice rules — specific structural directives (sentence length, lead placement, hedge prohibition)
- Brand facts — concrete claims about the product that must be stated accurately
- Format rules — structural patterns that are required or prohibited
Each component is a list of named, specific rules. The guide can still exist as a reference document for human onboarding. The constitution is what you give AI.
Copper Sun's how-it-works page covers how brand context is structured and injected into content generation — the constitutional model is the technical architecture, not just a prompting convention.
The gap between guide and constitution in practice
Consider a B2B software brand whose guide says: "We speak to sophisticated buyers who don't need hand-holding. We respect their intelligence." A reasonable human reads that and writes something like: "The system generates three-tier access controls automatically."
An AI working from the same description might produce: "Our innovative platform intelligently leverages advanced access control capabilities to empower your team with robust, enterprise-grade security solutions." Everything in that sentence traces back to plausible interpretations of "sophisticated" and "innovative" — neither of which appears in the source guidance. The AI did not fail; the guidance did not constrain.
The constitutional version of the same rule might be: Sophistication rule: State what the product does, not what it is like. Replace "robust security" with the specific feature. No innovation claims.
That rule eliminates the failure mode. The output becomes "The system generates three-tier access controls automatically." Not because the AI judged correctly, but because the rule made judgment unnecessary.
This is what the brand voice guide post covers in more depth — the specific properties of voice guidance that carry through AI generation and the ones that do not. And for teams building persistent brand context that applies across every project, the brand memory layer post covers how to structure constitutional rules for long-term retrieval.
Frequently Asked Questions
Can I use my existing brand guide and just add rules on top?
Yes, with qualification. The brand guide can stay as a human reference document. What you extract from it is a set of specific, named rules — not a summary or condensed version of the guide, but the directives that can actually constrain AI output. That extraction often reveals that a substantial guide contains very few rules that are specific enough to use. The process of writing the constitution is the clarifying work, not the compression.
How many rules does a brand constitution need?
Enough to cover the failure modes that actually appear in your AI content, and no more. Most teams find that a working set is between 8 and 20 rules. Fewer than 8 usually leaves significant gaps. More than 20 raises questions about which rules are load-bearing and which are aspirational — and aspirational rules do not improve output.
What is the difference between a terminology ban and a voice rule?
Terminology bans are word-level: specific words or phrases that should not appear. Voice rules are structural: patterns governing how sentences are built, how arguments are sequenced, or how claims are stated. Both are directive. The distinction matters because violations look different — a terminology violation is detectable by string match, a voice rule violation requires reading the sentence structure.
Does the constitutional model work for all content types?
The model works wherever you have a consistent brand voice to protect. It is most powerful for high-volume content — email sequences, product copy, social content — where the cost of description-driven drift accumulates quickly. It is less urgent for one-off pieces where a writer can apply judgment manually. The constitutional model does not replace editorial judgment; it makes judgment unnecessary for the rules you have already decided.