A Content Brief Template That Gets Specific Drafts From AI Writers

A content brief template for AI writing tools avoids generic output when it supplies what the model has no way to infer: who is asking, the one claim the page has to prove, and the evidence that proves it. In our view an outline and a keyword list are the inputs an AI writer needs least, because structure is the part a model can draft from the topic alone. A brief should spend its length on the inputs instead, and seven fields cover them. Each field below is tied to what Anthropic, OpenAI and Google publish about prompting their own models, and to the self-assessment questions Google asks site owners to put to their pages.

The brief to copy and fill in

Paste the filled brief into the same prompt as your drafting instruction. The guidance on generative AI content from Google Search Central describes the mechanism you are working against: "Keep in mind that generative models don't retrieve facts, but predict a likely sequence of words based on their training data." A field left blank gets filled the same way, with the likeliest sentence available.

SEARCH INTENT
Query the page answers:
What the searcher is trying to decide or do:
What they already read and found too thin:

READER
Role, situation and what they already know:
The objection they arrive with:

CLAIM TO PROVE
One sentence a reader could disagree with:

SUPPLIED EVIDENCE (use only this; write [NEED SOURCE] where it runs out)
First-party numbers, each with its date and sample:
Primary documents, pasted in full:
Fixed facts (product names, prices, definitions):

EXAMPLES
Two or three real cases, each with a detail an outsider would not know:

VOICE SAMPLES
Three to five passages we published, different in subject and structure:

PHRASING
Write like this, and why:
Leave out:

What does each field change in the draft?

A content brief template for AI writing gets its specificity from the audience, the evidence, the examples, the search intent and the claim, and each field should change something you can point to in the draft.

The vendors' documentation explains why the first five rows work. Anthropic tells prompt writers: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows." The search intent and reader fields are that missing context. OpenAI lists Context as one of four parts of a developer message, after Identity, Instructions and Examples, and describes it this way: "Give the model any additional information it might need to generate a response, like private/proprietary data outside its training data, or any other data you know will be particularly relevant." Google AI for Developers makes the same point for Gemini: "You can include instructions and information in a prompt that the model needs to solve a problem, instead of assuming that the model has all of the required information."

The evidence field carries two instructions on top of the material: use only what is supplied, and mark the gaps. Both follow the hallucination guidance from Anthropic, which recommends giving the model explicit permission to admit uncertainty and adds: "Explicitly instruct Claude to only use information from provided documents and not its general knowledge." A marker is quicker to resolve in review than a plausible invented number is to catch. The same page states the ceiling plainly: "Remember, while these techniques significantly reduce hallucinations, they don't eliminate them entirely." The fixed-facts line belongs in this field too, so every draft names products and prices identically; our entity consistency audit covers why that agreement matters once assistants start describing your brand.

How should the brief be structured so the draft keeps your voice?

To structure a content brief for AI drafting in a unique voice, show the voice with samples, give every ban a replacement, and label the sections so their order can change with the model. Adjectives such as confident or approachable fit almost any brand, so they give a model little to match. Published passages do more. Anthropic's documentation says, "Examples are one of the most reliable ways to steer Claude's output format, tone, and structure." It recommends three to five of them, relevant to the task and varied enough that the model does not pick up unintended patterns. Google's Gemini documentation goes further ("We recommend to always include few-shot examples in your prompts.") and warns that with too many examples the model may start to overfit its response to them. Choose passages that differ in subject and structure and share only the cadence you want copied.

A banned list belongs in the phrasing field, and a list alone does half the job. Anthropic's formatting advice carries the heading "Tell Claude what to do instead of what not to do", and its guidance on context says that explaining why an instruction matters can help the model deliver more targeted responses. Google's constraint guidance for Gemini allows both directions: "You can tell the model what to do and not to do." We put the positive instruction and its reason first, then a short leave-out list, because a ban with no replacement leaves the model to reach for the next most likely phrase.

Order is where the vendors' advice splits most clearly. Anthropic advises placing long documents (it gives 20k+ tokens as the threshold) near the top of the prompt, above the query, instructions and examples. Google's guidance for Gemini 3 says that when a prompt carries a large amount of context, all of it should come first, with specific instructions or questions at the very end, though essential constraints and output format requirements belong in the system instruction or at the very beginning of the prompt. OpenAI says of its Context section: "This content is usually best positioned near the end of your prompt, as you may include different context for different generation requests." OpenAI also notes that Markdown headers and XML tags help a model see the logical boundaries of a prompt, which is why every field in the template has its own heading: a labeled block can move without being rewritten.

One filled brief and the opening it is built to produce

Everything in this example is hypothetical. The HVAC installer, its job log, its figures and both openings were written for this article to show the mechanism, and neither opening is a logged model run. Test the pair on your own model before taking our word for the difference.

SEARCH INTENT
Query: heat pump vs gas furnace cost in a cold climate
Deciding between two installer quotes; wants a number and a rule.
Found too thin: comparison pages that list pros and cons for every home.

READER
Homeowner, house built before 1980, gas line already in.
Objection: a heat pump will run up the electric bill in January.

CLAIM TO PROVE
In a leaky older house, air sealing changes the answer more than
the choice of equipment does.

SUPPLIED EVIDENCE (use only this; write [NEED SOURCE] where it runs out)
- 2025 job log: 61 heat pump installs; 38 needed air sealing first
  to hit the sizing target. (hypothetical figures)
- Fixed facts: we quote both systems on every visit.

EXAMPLES
- 1962 split-level: sealing the attic hatch and rim joist dropped
  the quote by one equipment size. (hypothetical)

VOICE SAMPLES
(three passages from our own published posts, pasted in full)

PHRASING
Write to one homeowner as "you" and lead each section with the number,
because readers are comparing our quote with a rival's and discount
anything that sounds like a pitch.
Leave out: "look no further", "in the world of", "whether you're a".

A bare instruction to write about heat pumps and gas furnaces invites an opening like this one:

Choosing between a heat pump and a gas furnace is one of the biggest decisions a homeowner can make. Both systems have advantages and drawbacks, and the right choice depends on your climate, your budget and your energy goals.

The filled brief gives the model enough to open this way:

If your house was built before 1980 and you have two quotes on the table, seal the attic hatch and the rim joist before you choose equipment. In our 2025 job log, 38 of 61 heat pump installs needed air sealing first to hit the sizing target, and on one 1962 split-level the sealing work dropped the quote by one equipment size. NEED SOURCE: January operating cost at current utility rates.

Each specific in the second opening traces to a field: the reader's situation, the claim, one line of evidence, one example and one flagged gap. The first could sit on any HVAC site, which is a fair working definition of generic.

Which of Google's people-first questions does the brief answer?

The self-assessment questions that Google Search Central publishes for helpful, people-first content work as an acceptance test for the brief before drafting starts. Four of them map to fields:

• Supplied evidence: "Does the content provide original information, reporting, research, or analysis?"
• Examples: "Does your content clearly demonstrate first-hand expertise and a depth of knowledge (for example, expertise that comes from having actually used a product or service, or visiting a place)?"
• Search intent and reader: "After reading your content, will someone leave feeling they've learned enough about a topic to help achieve their goal?"
• An empty evidence field: "Are you mainly summarizing what others have to say without adding much value?"

When the evidence and examples fields stay empty after an honest attempt, that is a finding about the page, and prompt wording will not repair it. Google's generative AI guidance places the tool accordingly: "Generative AI can be particularly useful when researching a topic, and to add structure to original content." The brief is how the original content gets in. The same page sets the review duty ("It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing."), and a brief shortens that review by confining the draft to material you already vetted.

The policy line sits further out. The spam policies from Google Search Central define it: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users." For a description of the output to avoid, the generative AI guidance points to section 4.6.6 of the Search Quality Rater Guidelines, on "main content created with little to no effort, little to no originality, and little to no added value", while noting that raters' ratings "don't directly influence ranking". If the page is also meant to earn mentions in AI answers, our guides on how to get cited by ChatGPT, Perplexity and AI Overviews and how often AI answers cite sources set out what the published evidence supports.

Where does a filled brief still go wrong?

A filled brief takes longer to prepare than a one-line prompt, and most of that time goes to finding evidence. It also has limits. An error in the brief arrives in the draft stated with confidence, so the fact-check has to cover the brief's own numbers. Voice samples that share a subject can pull that subject into the draft, which is the kind of unintended pattern Anthropic's documentation warns about. A brief also cannot produce first-hand material a team does not have, and in that case the honest options are to gather it or to leave the page unwritten.

Anthropic's test for a prompt is to show it to a colleague with minimal context on the task and ask them to follow it. A version of that works on a brief: hand the filled brief to someone outside the project and ask what the article will argue and what proves it. Their answer tells you what the model will have to work with. Once the pages are live, AI share of voice is the measure we would watch for whether assistants pick them up. Further resources: the AI Visibility Checker as a starting point for your own site, and our AI search optimization service for teams that want this work run for them.

Sources

• Anthropic: Prompting best practices, Claude documentation
• Anthropic: Reduce hallucinations, Claude documentation
• OpenAI: Prompt engineering, OpenAI API documentation
• Google AI for Developers: Prompt design strategies, Gemini API, last updated September 17, 2026
• Google Search Central: Creating Helpful, Reliable, People-First Content, last updated October 5, 2026
• Google Search Central: Google Search's Guidance on Generative AI Content on Your Website, last updated October 1, 2026
• Google Search Central: Spam Policies for Google Web Search, last updated August 28, 2026