Brief templates for AI writers: the moat is the brief, not the model.

The brief is the moat.

Every team I have seen build an AI content pipeline starts by tuning the writer prompt. They iterate on system messages, swap models, test temperature, debate Claude versus GPT versus Gemini. After three months the output is still mediocre.

The teams that ship genuinely useful content do something different. They put 80% of their effort into the brief — the structured input that goes into the writer — and 20% into the model. The brief is where the strategy lives. The model is interchangeable infrastructure.

This is the brief template I run, why each field exists, and the compounding effect that makes a brief library defensible.

Why the brief and not the model

Open ChatGPT, Claude, and Gemini in three tabs. Give all three the same prompt: “Write a 1,500-word blog post about agentic SEO.”

The outputs will be different in style and slightly different in structure. None of them will be good. They will all be average — high-quality average, but average for the topic.

Now give all three the same 1-page structured brief: intent, audience role, exact entities, evidence requirements, tone exemplars, acceptance criteria. The outputs converge. They are recognizably good or recognizably need a second pass, regardless of the model.

The model contributes voice and minor mechanical quality. The brief contributes strategy. If the brief is empty, the model fills it with the average of the web. If the brief is specific, the model executes against your specificity.

This is why model wars are mostly noise for content pipelines. See best AI SEO tools 2026 for the broader stack picture.

The 8 fields a working brief has

After running this on 8 boutique pipelines, the same 8 fields appear in every brief that ships work that ranks.

1. Intent

One word: informational, commercial, transactional, or navigational. This is the single most consequential field. An informational brief that gets written as a commercial sales pitch ranks for nothing. Get this wrong and the rest of the brief is wasted.

2. Target query

The exact phrase you want to rank for. Not a topic, not a theme — the phrase a buyer types into Google or asks Perplexity. This becomes the H1 in slightly varied form, the H2 in question form (for AI citation, see how to get cited by ChatGPT), and the anchor text in internal links pointing in.

3. Audience operator role

Not “marketers” or “founders.” Specific: “Head of Marketing at a Series B B2B SaaS, post-PMF, manages 2 writers.” The writer agent will calibrate vocabulary, examples, and the level of explanation based on this field. Vague here = generic everywhere.

Include three sub-fields:

• Role. Title, company stage, team size.
• Knows already. What you do not have to explain.
• Pain. The specific business problem this article is supposed to relieve.

4. Entities to cover

Five to twelve named entities the article must mention. Pull these from Surfer or Frase based on top-10 SERP coverage, then prune to the ones that genuinely belong.

Examples for an article on agentic SEO:

• agentic SEO
• writer agent / editor agent
• brief library
• eval rubric
• topical authority
• content unit economics

This field is what gives the article entity richness — a major GEO citation signal (see entity-based SEO).

5. Evidence requirements

Specific demands for receipts. Examples:

• 3+ verifiable numbers per H2
• 1 named case study (anonymized OK)
• 2+ outbound citations to primary sources
• 1 internal link to a pillar page
• 0 unsourced superlatives

This field is what stops the writer from hallucinating. It is also what stops the writer from filing a vague essay full of “studies suggest” and “experts agree.”

6. Structural skeleton

The H2 list, with word targets per section. Not a full outline — a skeleton. Example:

• H2: what changed in 2025 (250 words)
• H2: the new cost stack (400 words)
• H2: what the brief library actually contains (350 words)
• H2: the eval loop math (300 words)

The writer fills in the sections. The skeleton stops the writer from inventing its own structure, which is where most off-topic drift originates.

7. Tone exemplars

Three to five paragraphs from existing content the writer should match. Not a description (“punchy, operator voice, short sentences”). Actual paragraphs the writer reads and imitates.

For my pipeline:

tone_exemplars:

• /writing/generative-engine-optimization.md (paragraphs 2-4)
• /writing/programmatic-seo-architecture.md (paragraphs 1-3)

Tone descriptions get interpreted differently every run. Tone exemplars converge fast. This is the single field where I trust no abstraction.

8. Acceptance criteria

Programmatic gates the editor agent or the eval rubric checks before the post is allowed to ship.

• Length range (e.g., 1800 to 2400 words)
• Banned-words scan passes
• Entity coverage (e.g., 6 of 6 required entities mentioned)
• Passes eval rubric v3

See eval loops for AI content for what the rubric contains.

If acceptance criteria are not in the brief, you have a creative writing prompt, not a content pipeline.

Why briefs compound

The first brief I wrote took 4 hours. The hundredth took 12 minutes. The reason is not skill — it is library effects.

Three assets accumulate.

The entity graph. Once you have written 30 briefs in a topic area, you have a deduplicated list of every entity that matters in the space, with notes on which articles each one is for. Filling field 4 (entities) collapses from “search the SERP, extract, prune” to “lift from the graph.”

The tone exemplars. Field 7 reuses across briefs in the same domain. After 30 briefs the exemplar library is stable.

The acceptance criteria templates. Field 8 is structurally identical across most posts. Length and entity coverage change; the rest is constant.

The first brief in a new client domain is still 4 hours of work, because the library has to be built. The second brief in the same domain is 90 minutes. The tenth is 25 minutes. The thirtieth is 12.

This is the difference between a content shop billing $1,200 per post and one billing $40 per post. The cheap shop has a library. The expensive one writes every brief from scratch.

What goes wrong without a brief

Three failure modes I have seen repeatedly.

Hallucination. Writer invents statistics, sources, case studies. Almost always traces to a missing or vague “evidence requirements” field. The writer fills the gap with plausible-sounding noise.

Off-topic drift. Writer wanders into adjacent topics, makes the article read as a generic explainer. Traces to a missing structural skeleton or vague H2s.

Tone mismatch. Writer produces text in a “thought-leader” voice when the brand voice is operator-direct. Traces to tone descriptions instead of exemplars.

Every one of these is a brief problem, not a model problem. Switching from GPT-5 to Claude does not fix it. Tightening the brief does, in one iteration.

The single-screen rule

A brief should fit on one screen. If you cannot see the whole thing without scrolling, you are writing too much.

The constraint is intentional. A brief that fits on one screen is one you can edit fast, hand to a collaborator, audit at a glance, and version cleanly in a library. A brief that runs to three pages is content with extra steps — you should have just written the post yourself.

The 8 fields above fit on one screen when you keep each field tight. Intent: one word. Target query: one phrase. Audience: three lines. Entities: a bullet list. Evidence: 4 to 6 bullets. Structure: 4 to 6 H2s with word counts. Tone: 3 to 5 file paths. Acceptance: a small object.

Where the brief stops and the writer starts

The brief is constraints, not content.

If you find yourself drafting actual sentences inside the brief, two things have gone wrong. First, the brief is leaking into the deliverable, which means you should just write the post. Second, you do not trust the writer agent enough — which usually means the eval rubric is too weak.

The right relationship is: the brief sets a tight perimeter, the writer fills the interior, the editor checks the perimeter held. If any of those three stages is sloppy, the others compensate badly. Keep them clean and separate. See agentic SEO cost economics for what this separation does to unit cost.

What to do tomorrow

If you are starting from zero, the first move is small.

1. Open the brief template at the top of this post.
2. Pick one upcoming article. Fill in all 8 fields. Time it.
3. Run the brief through whichever writer you are currently using. Compare to your last freelance draft.
4. Note what is missing or wrong in the output. Add an evidence requirement or tighten an exemplar to fix it.

The brief gets faster every time. The writer gets better every time. After 10 briefs you stop wondering whether AI content can work. After 30 you wonder how anyone ships without a library.