The AI Content Brief Template That Prevents Generic Drafts

Most people blame the model when AI output reads like a press release written by a committee. The model is rarely the problem. A brief that says “write a 1,500-word article about email deliverability for B2B SaaS marketers, professional but friendly tone” contains almost no information that a language model can’t already guess, so it does exactly what you’d expect: it returns the statistical centre of everything ever written about email deliverability. Generic in, generic out.

The fix is a brief structured around the things a model cannot know. Your customer’s actual objections. The number your client reported last quarter. The specific argument your competitor got wrong. This page gives you a template built around that principle, with a worked example and the rules for when to skip parts of it.

What a generic draft actually looks like, and why

Run this prompt through Claude, ChatGPT or Gemini: “Write an 800-word blog post about why content teams should audit their content. Audience: in-house marketers. Tone: practical, no fluff.”

You’ll get something with the following properties, near enough every time:

• An opening paragraph beginning “In today’s fast-paced digital landscape” or a lightly disguised variant
• Three to five H2s with near-identical parallel structure (“Identify Underperforming Content”, “Set Clear Goals”, “Measure Your Results”)
• Advice that is true, safe and applies equally to a 12-person SaaS team and a national charity
• Zero numbers that aren’t round
• A closing paragraph that restates the H2s

None of that is a failure of the model. It’s a faithful response to a prompt that specified only topic, length and audience. Every remaining decision got filled in by the prior. If you want output that sits outside the prior, the brief has to push it there.

The practical test I use: could a competitor’s content team have written this brief? If yes, the draft will be generic no matter which model you run it through. A good brief should be slightly embarrassing to share publicly because it contains your actual positioning, your actual sales objections and your actual numbers.

The template

Copy this into Notion, Google Docs or wherever your briefs live. It’s designed to be filled out in 15 to 25 minutes for a standard supporting page, which is roughly what a decent traditional brief takes anyway.

## 1. The one question
What single question does this page answer?
(One sentence. If you need two, you have two pages.)

## 2. Who is asking, and what they've already tried
Role, seniority, company size:
What they already know (so we don't explain it):
What they've already tried that didn't work:
What they're afraid of / what the cost of getting this wrong is:

## 3. Our actual position
What do we believe about this that isn't consensus?
What common advice on this topic do we think is wrong, and why?
What are we NOT claiming? (guards against overreach)

## 4. Proof inventory
Numbers we can cite (source + date):
Customer examples / anonymised stories:
Screenshots, tool outputs or data we own:
Named tools, products or standards to reference:

## 5. Structural constraints
Primary keyword:
Word count range:
Required sections:
Internal links (URL + why a reader would click it there):
CTA and what happens after the click:

## 6. Anti-brief: what a bad version looks like
Three phrases, framings or structures this draft must not use:
The obvious angle we're deliberately avoiding:

## 7. Voice samples
Two or three links to our own published pages that sound right.
(Paste 300–500 words of the best one directly into the prompt.)

Sections 3, 4 and 6 do most of the work. They’re also the three that get skipped when people are in a hurry, which is why so many teams conclude AI writing doesn’t work for them.

Section 6 is the one nobody writes, and it matters most

The anti-brief is the highest-leverage part of this template. Models are considerably better at avoiding a named pattern than at inventing an unnamed one. “Be original” does nothing. “Do not open with a statistic about how much content is published per day, do not use the phrase ‘in an era of’, and do not structure this as a numbered listicle” does a lot.

For a page about content audits, a realistic anti-brief might read:

Must not use: “content audit” framed as a spreadsheet exercise; the phrase “low-hanging fruit”; any framing that implies you need a dedicated tool to start. Obvious angle we’re avoiding: the step-by-step “how to run an audit in 10 steps” structure. Twelve competitors rank for that. We’re writing about the decision of what to do with the pages you find, which nobody covers properly.

That second line changes the entire draft. It’s also the sort of thinking that belongs upstream of the brief, in the planning work covered on the AI-assisted content strategy and planning pillar. If you don’t have a clear view of where a page sits in the wider programme, the brief will inherit that vagueness.

A worked example

Here’s a filled-in brief from a real-ish scenario: a UK B2B software company, roughly £6m ARR, four-person marketing team, writing a supporting page under a pillar about procurement software.

1. The one question: Why do purchase order approvals still take eight days at companies that already bought approval software?

2. Who is asking: Finance operations managers at UK companies with 200 to 800 staff. They already have Coupa or Proactis. They’ve already tried shortening the approval chain and it caused a compliance flag at audit. They’re afraid of being the person who loosened controls before something went wrong.

3. Our actual position: The bottleneck isn’t the approval chain length, it’s that approvers can’t tell from the notification whether a PO is routine or exceptional, so they treat everything as exceptional. Common advice says reduce approval steps. We think that’s wrong and mildly dangerous. We’re not claiming that approval chains never need trimming.

4. Proof inventory: Our own 2026 customer data: median approval time fell from 8.2 days to 2.9 days across 34 accounts after enabling risk-scored notifications, with no change in chain length. Named tools: Coupa, Proactis, SAP Ariba, Xero for the SME end. Anonymised story: a housing association where 71% of POs were under £500 and routed identically to £40k contracts.

5. Structural constraints: Primary keyword “purchase order approval time”. 1,400 to 1,800 words. Must include a section on audit defensibility. Link to the pillar page and to the risk-scoring feature page. CTA: book a 20-minute workflow review, not a demo.

6. Anti-brief: No “digital transformation”. Don’t frame this as a paperless-office story. Don’t open with a statistic about manual processes. Avoid the obvious “5 ways to speed up PO approvals” structure.

7. Voice: Paste in our “Why your spend under management number is probably wrong” post.

Feed that into Claude with an instruction to write from the brief and ask clarifying questions first, and the resulting draft has a spine. It has an argument the competition isn’t making, a number nobody else can cite and a clear thing it refuses to say. That’s the whole game.

Running the brief through a model without losing it

A few mechanics that make a difference in practice:

Put the brief in the system prompt or a project, not the chat turn. In Claude Projects or a ChatGPT custom GPT, load the brief plus your voice samples as persistent context. In the API, put them in the system prompt. Long chat threads drift; persistent context doesn’t. If you’re in a plain chat window, re-paste sections 3 and 6 before any significant revision request.

Ask for the outline first, and argue with it. Send the brief, ask for five candidate H2 structures, then reject four. This takes about three minutes and catches the generic structure before 1,500 words have been built on top of it. It also surfaces where your brief was thin: if all five outlines look the same, section 3 is the problem.

Never accept the first draft’s opening paragraph. Ask for six alternative openings, each taking a different entry point (a specific scene, a contradiction, a number, a question the reader has asked out loud, an objection, a flat statement of the position). One of the six is usually usable with light editing.

Keep a rejected-phrase list per client or brand. Mine currently runs to about 40 entries. It goes into every brief’s section 6 automatically. Adding “delve”, “landscape”, “leverage” as a verb, “it’s not just X, it’s Y” and “in today’s” to that list removes most of the tell.

When to skip half of this

Not every page deserves 25 minutes of briefing. A rough allocation that works for small teams:

The section that never gets skipped is 3. A page with no position doesn’t need AI to be generic, it was going to be generic anyway.

The measurable difference

Worth tracking whether this is actually working rather than assuming it. Three lightweight measures:

Edit distance. Count how many paragraphs of the AI draft survive to publication roughly intact. Briefed properly, expect 50 to 70% survival on a supporting page. Under 30% means the brief failed and you wrote the thing yourself, expensively.

Time from brief to publishable draft. Teams I’ve seen do this properly land at 90 minutes to two hours for a 1,500-word supporting page, including the briefing time. If you’re at four hours, the brief is doing too little work upstream.

The blind read. Once a month, put one AI-assisted page and one fully human page in front of someone who knows your brand and ask which is which. When they can’t reliably tell, your briefs are carrying their weight. When they nail it every time, read the pages they flagged and look for which brief section was empty.

There’s one failure mode worth watching for. Teams that get good at this start briefing so thoroughly that the brief contains the article, and at that point you’re writing twice. The brief should specify the argument, the proof and the prohibitions. It shouldn’t specify the sentences. If you find yourself drafting paragraphs inside section 3, close the brief and go write the page.