I've spent the better part of two years running client blog posts through ChatGPT, Claude and Gemini, and the pattern took me a long time to spot. When a draft came back flat, I blamed the model. I'd switch tools, try a new plugin, rewrite the prompt three times at eleven at night. What finally fixed it was dull: I stopped typing prompts and started handing the tool a proper brief, the same kind I used to send freelance writers.
The difference was immediate. Same model, same topic, and the draft came back with the right angle, the right reader in mind and none of the padding I'd been cutting out by hand. Since then I've kept a brief template in a pinned doc and filled it in before every single piece.
This document is that template, along with what I learned about which sections matter and which ones the model quietly ignores. Nothing here needs a paid tool. Fill in the nine sections, paste it in, and you get a first draft that needs editing instead of rescuing.
The short version A content brief for AI is a structured document you paste in before asking for a draft. It tells the model who the reader is, what the piece must achieve, what it must include, what it must leave out and how it should sound. A good one runs 300 to 600 words and replaces an hour of prompt tweaking with one complete set of instructions. |
Why a prompt is not a brief
Most people write prompts the way they'd message a colleague: "Write a 1,500-word blog on content briefs, make it engaging." The model then fills every gap with the most statistically average answer it has, which is exactly the thing you don't want on your site. Four things go wrong when the brief is missing.
It guesses the reader With no audience named, the draft aims at everyone and lands with no one. You get definitions a beginner needs sitting next to jargon only an expert would follow. | It repeats the internet In Ahrefs' July 2026 study of top-ranking pages, the team singled out one habit of weak AI drafts: restating common knowledge without adding anything the search results didn't already have. The brief is where you supply the new part. |
It invents the missing facts Ask for statistics without providing any and the model produces plausible numbers. Some are right. You won't know which ones without checking every single one. | It falls back on default AI voice Long tidy paragraphs, three-item lists, the same transition words. Without voice rules in the brief, every draft sounds like every other draft published that week. |
Ahrefs found that 74.2% of the 900,000 new pages it analysed in April 2025 contained some AI-generated text, and only 2.5% were fully AI-written. Most of the web is now a blend. What separates the blends that rank from the ones that don't is what the human put in.
The nine things to give AI before it writes
These are the sections I fill in every time. The third column is what I saw happen, repeatedly, when I skipped each one.
| Section | What to put in it | What happens if you skip it |
|---|---|---|
| 1. Title and angle | The headline you'd be happy to publish, plus one sentence on the specific take. "How to brief AI" is a topic. "Why prompts fail and briefs don't" is an angle. | You get the generic version of the topic. |
| 2. Reader profile | Their job or situation, what they already know, and what they want to do after reading. | Mixed-level content that satisfies nobody. |
| 3. Search intent | The main query, the questions people ask around it, and whether they want a how-to, a comparison or a definition. | The draft answers a different question than the one that ranks. |
| 4. Facts and sources | Statistics, quotes, product details and internal data, each with a source link. Mark anything the model must not loosely paraphrase. | Fabricated or outdated numbers. |
| 5. First-hand material | Your experience, test results, screenshots, customer conversations. This is the part no model can supply. | Content that reads like a summary of other content. |
| 6. Structure | Section headings in order, rough word count per section, and where tables or lists belong. | Paragraph walls under headings that mirror competitors. |
| 7. Voice and style | Sentence length, contractions yes or no, banned words, punctuation preferences, reading level, and two sample paragraphs in the right voice. | Default AI cadence that editors spend an hour removing. |
| 8. Exclusions | Competitor names to avoid, claims you can't back up, topics already covered elsewhere on your site. | Brand and legal problems that surface after publishing. |
| 9. Output format | Total length, heading levels, HTML or Markdown or plain text, and how many title options you want. | Reformatting work on your side. |
Sections 4 and 5 do most of the work. They're also the two people leave empty most often, because they require you to know something. If you only have ten minutes, spend them there.
The template
Copy this into a note, fill in every line, and paste the whole thing as the first message of a new chat. The last line is the most useful sentence in the whole document; I'll come back to why.
CONTENT BRIEF 1. WORKING TITLE: Angle (one sentence): 2. READER Who they are: What they already know: What they want to do after reading: 3. SEARCH INTENT Primary query: Related questions: Intent type (how-to / comparison / definition / list): 4. FACTS TO INCLUDE (each with a source link) - - Do not use any statistic that is not listed here. If a claim needs a number I have not supplied, write [NEEDS SOURCE]. 5. FIRST-HAND MATERIAL (paste notes, test results, quotes, observations) 6. STRUCTURE H2: approx. words: H2: approx. words: H2: approx. words: Use a table for: Use a list for: 7. VOICE AND STYLE Reading level: Sentence length: Contractions: yes / no Banned words and punctuation: Reference sample (two paragraphs in the right voice): 8. DO NOT INCLUDE - - 9. OUTPUT Total length: Format (HTML / Markdown / plain text): Title options: Meta description: yes / no Before drafting, list every section of this brief where the information is missing or unclear. Ask me, do not assume. |
That closing instruction changes the model's behaviour more than anything else in the brief. Without it, a tool will happily write around a gap. With it, you get a short list of questions first, usually the same ones an editor would have asked.
A filled-in example
A trimmed version of a brief I'd write for a small accounting firm's blog, to show the level of detail that works. The real one is about twice this length once the facts are in.
| Working title | How to Prepare for a Tax Enquiry as a Freelancer |
| Angle | What the letter actually asks for, based on the enquiry letters our clients receive, not a generic "stay calm" piece. |
| Reader | UK sole trader, first enquiry, never audited, wants a checklist they can work through this weekend. |
| Search intent | "freelancer tax enquiry checklist". How-to. Related: how long does it take, what records do they ask for. |
| Facts | Three points from HMRC's published compliance-check guidance, each with the page link. Our own figure for average time to close, with the internal report attached. |
| First-hand | Two anonymised client stories from the partner's notes, pasted in full. |
| Structure | Five H2s, 250 words each, a table of documents to gather, a numbered list for the first week. |
| Voice | Plain English, short sentences, contractions allowed, no exclamation marks, no "navigate", no "robust". Two sample paragraphs from our best-performing post attached. |
| Exclusions | No penalty figures (they change), no naming of software vendors, nothing on limited companies. |
| Output | 1,400 words, Markdown, three title options, meta description included. |
Where to store the brief in each tool
Every major assistant now has a place to keep standing instructions, so you don't paste the same voice rules every time. The per-piece brief still changes with each article. The reusable parts can live inside the tool.
| Tool | Where the standing rules live | Worth knowing |
|---|---|---|
| ChatGPT | Projects hold shared files and instructions across many chats. Custom GPTs package a persona you can hand to a teammate. | Custom instructions have a character cap (1,500 on free plans and 5,000 on paid at the time of writing), so keep standing rules tight and put long material in project files. |
| Claude | Projects take project-level instructions plus uploaded documents. Styles control tone separately from the instructions. | Good fit for document-heavy work. You can drop in a full style guide and several sample posts and reference them by name in the brief. |
| Gemini | Gems store persistent instructions and can link a Google Drive document as knowledge that updates live. | The instruction field is short (recent comparisons put it around 4,000 characters), so link to the long material instead of pasting it. |
| Any tool | The per-piece brief goes in as the first message, every time. | Don't rely on memory features for facts. They store preferences about you, not sources for your article. |
What changes in the draft
This is what I saw shift once I stopped prompting and started briefing, across the same handful of models and the same kinds of topics.
| What you're looking at | One-line prompt | Full brief |
|---|---|---|
| Who the draft is aimed at | Everyone, so nobody | The one reader named in section 2 |
| Statistics | Unsourced, some invented | Only the ones supplied, with links intact |
| Structure | Headings that mirror the top results | The sections specified, with tables where asked |
| Voice | Default AI cadence | Close to the reference sample, light edits |
| Anything new in it | Rarely | Whatever went into section 5 |
| Editing job | Rewrite | Trim and verify |
That last row matters for search as much as for your afternoon. In Ahrefs' 2026 analysis of top-10 results, pages with low or moderate levels of AI-detected text earned two to three times the impressions of pages that were mostly AI. The researchers' read was not that Google punishes AI, but that quality tends to fall as the AI share rises. A brief with real facts and first-hand material keeps the quality up while you still use the tool.
Mistakes that undo a good brief
Briefing the topic, not the angle "Content briefs" is a topic. The model already knows what the top ten results say about it. Your angle is the one sentence that tells it what those pages missed. | Pasting the whole style guide A 4,000-word brand document buries the five rules that matter. Pull the rules out, list them, and add two sample paragraphs instead. |
Leaving section 4 empty An empty facts section is an invitation to invent. If you have no numbers yet, say so in the brief and tell the model to leave placeholders. | Dropping the "ask, do not assume" line It feels like a formality. It's the line that turns silent guessing into a short list of questions you can answer in two minutes. |
Never updating the template Keep a running list of things your editor keeps fixing and add a rule for each one. Mine grew from six style lines to about twenty over a year, and the drafts improved in step. | Asking for volume instead of value Google's spam policies name scaled content abuse, meaning many pages produced mainly to manipulate rankings, however they were made. A brief per page is the opposite of that pattern. |
Pre-paste checklist
Run through this before the brief goes into the chat. It takes about a minute and catches the gaps I used to notice only after reading the draft.
☐ One named reader, not a demographic
☐ Every statistic in section 4 has a link
☐ At least one thing in section 5 that isn't on the first page of results
☐ The banned-words list matches what your editor flagged last time
☐ Structure says where the table and the list go
☐ The final line asks the model to flag gaps before writing
Final verdictAfter a lot of hours with these tools, I've stopped thinking of the brief as prep work and started treating it as the actual job. Drafting is now the fast part. Whether a piece ranks or sinks is decided almost entirely before the model writes a word: did I name a real reader, did I hand over facts with sources, and did I put something in the brief that the model could not have found on its own. Two caveats. Filling in nine sections takes fifteen to twenty minutes the first few times, which feels slow if you're used to typing one sentence and pressing enter. It gets quicker, and the reusable sections mean repeat topics are mostly done before you start. Second, a brief does not remove editing. It changes editing from rewriting to checking. If you take one thing from this article, take the last line of the template. Asking the model to list what it doesn't know before it starts has caught more problems for me than any other single instruction. Copy the template, run one piece through it this week, and put the draft next to your last unbriefed one. The gap makes the argument better than I can. |
