What not to put in your AI marketing brief

The standard advice about AI briefs is to provide more context. More background on the company, more detail about the audience, more description of the brand's values and voice. The intuition is that more input produces better output.

The research says otherwise. A 2023 paper from Shi et al., published at ICML, found that adding irrelevant context to an AI task prompt dramatically decreases performance — even on tasks the model can otherwise solve correctly. The finding: irrelevant context degrades output even when the model has all the information it needs to succeed.

For marketing briefs, this finding has a specific implication: the content you add "just in case" is not neutral. It actively makes the output worse.

Why irrelevant context hurts

The mechanism is attention. AI models process context by attending to the parts most relevant to the task. When the context contains irrelevant content — information that is true but not directly useful to what is being generated — the model distributes attention across a larger context window. Relevant constraints receive proportionally less weight.

For a marketing brief, the practical effect is predictable: voice rules that are surrounded by company history, mission statements, and aspirational language compete with that context for attention. The output satisfies the whole brief, including the parts that should not have been there. The result is content that sounds broadly appropriate but does not follow the specific rules you actually care about.

A tight brief, containing only the constraints that apply to this generation, consistently outperforms a longer one.

What to leave out

Company history and background. The year the company was founded, the founding story, the growth trajectory, the current team size. None of these constrain what the AI writes. They are relevant to onboarding a human employee; they are not relevant to generating a blog post, email, or social update that follows your voice rules. Leave them out.

Mission and values statements. "We believe that every organization deserves access to better AI" is not a content rule. It is a true statement that does not tell the AI what to do differently. If your values imply content rules, derive the rules and include only the rules. The values themselves do not help.

Vague tone descriptors. "Professional but approachable." "Expert but accessible." "Confident but humble." These are balanced pairs that describe a feeling rather than a behavior. A voice rule like "Hedge rule: Replace 'may help' with 'reduces'" is specific. "Confident but humble" requires interpretation the AI will supply from its training distribution, not from your brand.

Aspirational brand positioning. "We are the leading provider in our category." "We are at the forefront of AI-powered marketing." These are self-characterizations that do not constrain content. If you have a specific, verifiable claim you want used, include the exact claim. If you are including positioning because it feels like it should be in the brief, leave it out.

Approval-seeking caveats. "Please feel free to adjust as needed." "Let me know if you need any clarification." These are human conversational conventions that function as noise in a brief. They do not change what the AI generates and they consume tokens that could be used for constraints that do.

Anything that starts with "we believe," "we strive," or "we value." These constructions describe aspirations rather than behaviors. An aspiration is not a constraint. If you believe in plain language, specify the plain-language rule. If you strive for authenticity, define what authentic means in your content and rule it accordingly.

The attention test

For anything you are considering including in a brief, apply the attention test: if the model paid full attention to this and nothing else, would the resulting content follow the rule I actually care about?

A terminology ban passes this test: full attention to "do not use leverage, ecosystem, scalable" produces content without those words. "We believe in clear, accessible language" fails this test: full attention to that sentence produces content that sounds broadly accessible, but may still use the specific words you are trying to avoid.

The test filters for constraining versus descriptive content. Include what constrains; leave out what only describes.

What a brief contains instead

A brief that has been through the attention test contains only:

1. Named voice rules — specific sentence-level behaviors
2. Terminology bans — explicit words and phrases to avoid
3. Required terms — words and phrases that must appear
4. Brand facts — exact phrasing for claims that must be accurate
5. Audience rules — what this audience knows and doesn't need explained
6. Format rules — structural patterns that apply to this output

Each of these is directly constraining. None of them are descriptive, aspirational, or contextual in ways that dilute the attention the constraints need.

The structured brand brief template covers each section in detail. The brand constitution post covers the underlying model — why named directives outperform descriptive guidance at every length.

Copper Sun's approach

Copper Sun's brand context system stores brief content as structured fields rather than as a single block of text. Voice rules, terminology bans, and brand facts are each stored separately and injected in a format that makes their constraining role explicit. This is the architectural version of the attention test: each constraint is labeled and positioned to receive full model attention rather than competing with surrounding text.

Frequently Asked Questions

How long should an AI marketing brief be?

Long enough to cover the constraints that apply, and no longer. A brief for a blog post might need 15 to 20 named rules. A brief for a social post might need 8 to 10. The correct length is determined by what actually constrains output, not by what feels thorough. If removing a section from your brief would not change the output, that section should be removed.

What about context that is technically relevant but borderline?

When in doubt, apply the attention test. If you are uncertain whether something is relevant, ask: if the model attended only to this, would it produce a behavior I actually want? If yes, include it. If the answer is "it would produce adjacent behavior that sort of relates to what I want," leave it out and specify the behavior you want directly.

Can context ever be too sparse?

Yes. A brief with only one or two constraints gives the model too much latitude — the output defaults to training distribution patterns rather than brand-specific patterns. The goal is not minimalism; it is precision. Every constraint that belongs should be included. Every constraint that does not belong should be excluded. The brief should be exactly as long as it needs to be.

Does this research apply to all AI systems?

The Shi et al. finding was tested across multiple reasoning and language tasks. The mechanism — attention dilution from irrelevant context — is a structural property of transformer-based models. Different models vary in how severely irrelevant context affects performance, but the direction of the effect is consistent: less relevant context produces better results than more irrelevant context. The practical guidance holds across the AI systems marketing teams are using today.