uprank
Sign inGet started

Guide

How to write a content brief for an AI blog writer

To write a content brief for an AI writer, name the reader and the one question the post answers, say what the search is really asking for, and pick an angle that adds something a summary would not. Then outline the sections with a purpose for each, attach the sources and list the claims the writer must not make, set the voice and the quality bar, and check the draft against the brief line by line. uprank writes that brief for you: a model drafts it with 18 fields from the research and your brand profile, a second model call reviews it, and it is rewritten up to 2 times if the review fails. You cannot edit or approve that brief, the draft is written from it at once, and no step checks the finished article against its outline, so a person still has to.

By Devino · Published · Updated

The method, step by step

Works for any post, whichever tool writes it.

  1. Step 1

    Name the reader and the one question the post answers

    Write down who is reading, what they are trying to get done and the single question the post has to answer. “Anyone interested in marketing” is not a reader. If you cannot state the question in one sentence, the draft will wander, and an AI writer will fill the gap with a generic overview.

  2. Step 2

    Say what the search is really asking for

    State the search intent in plain words: is the reader trying to learn something, compare options, fix a problem or buy? Then say what a complete answer looks like, such as steps, a comparison or a definition. That description is what the finished post gets measured against, by a person or by a check.

  3. Step 3

    Pick an angle that adds something a summary would not

    Decide what this post gives the reader that a summary of existing posts does not: a worked example, a framework, a comparison, your own numbers or a position you will defend. Write it down as a requirement. Without one, an AI writer defaults to restating what is already published.

  4. Step 4

    Outline the sections and say what each one is for

    List the headings in order and add one line per heading on its job: the question it answers and what the reader can do afterwards. Say what stays out as well. A purpose for every section is what lets you tell later whether a section earned its place.

  5. Step 5

    Attach the sources and fence the claims

    Name the sources the writer must use and the facts to check before publishing. Then list the claims the writer must not make: promises of results, statistics with no source, comparisons that name a competitor, anything legal or medical you have not cleared. Sources give the writer facts to use, and the do-not-say list stops claims you cannot back.

  6. Step 6

    Set the voice, the call to action and the quality bar

    Say how the post should sound, which words are off limits, what single action the reader is asked to take and what must be true for you to publish, such as “includes a worked example” or “every number has a source”. A bar written before the draft is easier to hold than a feeling after it.

  7. Step 7

    Test the brief, then check the draft against it

    Read the brief cold and ask whether two different writers would produce roughly the same article from it. Fix what is vague. When the draft arrives, check it against the brief line by line: each section present, each required example used, each banned claim absent. A brief nobody checks the draft against turns into a suggestion.

Checking the finished draft’s claims against their sources is its own job: how to fact-check AI content. How the brand’s voice and positioning reach the writer is how to keep AI content on brand.

What uprank does at each step

Each step of the method mapped to what the code does, including where it stops.

  • Name the reader and the one question the post answers. uprank does not ask you to write the brief. A model writes it as a step of a pipeline run, straight after the research step, from the topic, four fields of the topic’s Evidence Pack (search intent, reader problem, unique angle and current landscape), the site’s brand positioning and proof points, and the Evidence Pack’s numbered source excerpts. A site with no brand profile gets a brief written without brand input. The brief has fields for the intended reader, the reader problem and the primary question, and the model fills them in. In the app an Editor starts the run from a campaign page; over MCP, pipeline_run and article_generate start the same pipeline.
  • Say what the search is really asking for. The brief has a search intent field, and in a pipeline run that sentence is passed on to scoring. The model-judged Search intent dimension, 10% of the overall score and without a score gate of its own, is given it together with the start of the article and asked whether the article satisfies it. A version that is scored again over MCP, or revised with Apply selected or article_revise, is judged against the Evidence Pack’s search intent instead of the brief’s.
  • Pick an angle that adds something a summary would not. The brief has a unique angle and an information-gain requirement, and the instructions ask for an angle that reflects the brand’s proof points, not a generic summary. The draft is told to include one information-gain element and to record which one. Two checks follow: the “At least one information-gain element” integrity check, one of 13 hard gates, and a model-judged Information gain score that has to be ≥ 70. Neither compares the element the draft declares with the requirement written in the brief.
  • Outline the sections and say what each one is for. The brief has a section outline with a purpose for every heading, plus required examples. They reach the draft as text: the whole brief goes to the model that writes it, and its instructions say to follow the brief exactly. Nothing afterwards compares the finished headings or examples with the outline, so a draft can leave a section out or add one without any check noticing.
  • Attach the sources and fence the claims. The brief lists required sources and external references as numbers from the Evidence Pack’s numbered sources, claims to verify and claims to avoid. They are instructions to the writing model, not rules the checks apply. The checks that do run look at the finished article: each citation has to point to a source in the Evidence Pack, whether or not the brief listed it, and claims are taken from the article text and checked against the Evidence Pack’s sources. The brief’s own claims-to-avoid list is not checked against the draft. Forbidden claims and terms come from the brand profile.
  • Set the voice, the call to action and the quality bar. The brief carries tone instructions, a call to action, a visual requirement and a quality bar. These 4 fields are optional, so the model may leave them out, and they are text for the model. Brand fit is scored against the brand profile, not against the brief, and the 13 hard gates are the same for every article, so a quality bar written into the brief does not raise or lower any of them.
  • Test the brief, then check the draft against it. uprank tests the brief once, before the draft, with a second model call. It reviews the brief against a checklist (brand uniqueness, adds more than a summary, answers the reader’s real question, practical value, evidence sufficiency, tradeoffs, source currency, reputational and factual risk, and structure discipline) and returns pass or fail with the failing items and revision instructions. The reviewer sees the brief and the same context summary the writer had, not the source excerpts. On a fail the brief is rewritten, up to 2 times; if it still fails, the run stops with the error BRIEF_GATE_FAILED and no draft is written. A model reviewing a model’s work is not a person reading the brief, and no person is asked to before the draft starts. After the draft, nothing checks it against the brief line by line.

The full list of gates is on quality gates, and the product view of brief, draft and checks is AI blog writer vs AI chatbot for blog posts.

What is in the brief uprank writes

18 fields, shown read-only on the article’s Brief tab. 4 of them are optional, so the model may leave them out.

Working title
A working title for the post.
Intended reader
Who the post is for.
Search intent
What the person searching is trying to do.
Reader problem
The problem the reader arrives with.
Primary question
The one question the post has to answer.
Unique angle
What this post says that a generic summary would not. The instructions ask for an angle that reflects the brand’s proof points.
Information gain requirement
The element the post must add beyond a summary.
Section outline
The headings in order, each with its purpose.
Required examples
Examples the post should include.
Required sources
Which of the Evidence Pack’s numbered sources the post should draw on.
Internal link intents
Where the post should link to other pages on the site, in words.
External references
Which of the Evidence Pack’s numbered sources to point readers to.
Claims to verify
Claims the post will make that need checking.
Claims to avoid
Claims the post should not make.
Visual requirementoptional
What kind of image or visual the post needs.
CTAoptional
The call to action.
Tone instructionsoptional
How the post should sound.
Quality baroptional
What the finished post has to meet.

Ways to get a brief out of uprank

They differ in whether a draft follows, who can use them and where the brief ends up.

  • A pipeline run (campaign page, or pipeline_run)

    Researches the topic, or reuses an Evidence Pack the topic already has ready, then writes and critiques the brief, drafts, checks the claims and scores, without stopping for anyone to read the brief.

    Who can use it
    An Editor in the app, or an agent with the article:write and research:run scopes over MCP (pipeline_run).
    Writes a draft
    Yes, straight after the brief.
    The brief
    Stored with the article’s version and shown, read-only, on the article’s Brief tab.

  • article_generate (MCP)

    Starts the same pipeline for a topic opportunity. It writes a new brief for that run; it does not take one you generated earlier.

    Who can use it
    An agent with the article:write scope.
    Writes a draft
    Yes, straight after the brief.
    The brief
    Stored with the article’s version and shown, read-only, on the article’s Brief tab.

  • brief_generate (MCP)

    Writes and critiques a brief for a topic that already has an Evidence Pack, and stops. It refuses a topic with no Evidence Pack and does not research.

    Who can use it
    An agent with the article:write scope.
    Writes a draft
    No. It creates no article and no draft.
    The brief
    Only in that run’s output, which the run resource (run://workflow-runs/{id}, article:read scope) returns. It is attached to no article, and a later article_generate or pipeline_run writes its own brief instead of using it.

Where the brief is used once it exists

It steers the model. It is not a contract the checks hold the article to.

  • The first draft. The whole brief, as JSON, goes into the prompt of the model that writes the draft, next to the topic, the brand profile and the Evidence Pack sources.
  • The automatic revision. When a draft fails a gate, the pipeline writes revised versions from the same brief, not from the current text.
  • The Search intent score. In a pipeline run the brief’s search intent is what the judge is given (10% of the overall score).
  • The saved version. The brief is stored with each version that is drafted or revised, and the article’s Brief tab shows the current version’s brief, read-only. A version saved in the editor keeps the brief of the version it was edited from.
  • Later revisions. Apply selected on the Quality tab and article_revise write a new version from the brief stored on the version being revised.

What uprank will not do

  • Let you write or edit the brief. No screen and no MCP tool changes a brief, and the Brief tab is read-only. What you can change is what goes into it, the topic and the brand profile, and then run again for a new brief.
  • Ask a person to approve the brief. A pipeline run goes from the brief straight to the draft. Approval in uprank is of an article version, after the draft exists.
  • Stop after the brief inside a pipeline run. To get only a brief, call brief_generate over MCP; a later run does not use it.
  • Check the finished article against the brief. Only the search intent reaches a score, and only as a sentence the judge is given. The outline, required examples, required sources and claims to avoid are not compared with the draft, and the information-gain requirement is not compared with the element the draft declares.
  • Check the facts in the brief. Claims to verify is a list the writing model is given. Claims are checked on the finished article, against the Evidence Pack’s sources.
  • Turn the brief’s link intents into links. Link suggestions are made after the draft, internal ones from your site’s ingested pages and external ones from the Evidence Pack’s sources. Neither step reads the brief, and suggestions are never written into the article text.
  • Replace the review of the draft. The 13 hard gates apply to the finished article whatever the brief said, and a person still decides whether the article matches what the brief asked for.

Who approves a finished version is on the AI content approval workflow. This page makes no claim about how a brief changes how a post performs in search.

Common questions

How do you write a content brief for AI?
Name the reader and the one question the post answers, then say what the search is really asking for and pick an angle that adds something a summary would not. Outline the sections with a purpose for each, name the sources the writer must use and the claims it must not make, and set the voice, the call to action and the quality bar. Test the brief by asking whether two writers would produce the same article from it, and check the draft against it line by line. uprank writes a brief with 18 fields for you, but it does not check the finished article against that brief.
What should an AI content brief include?
At least the reader and their problem, the primary question, the search intent, the angle and what the post adds beyond a summary, a section outline with a purpose for each heading, examples, sources, claims to verify and to avoid, the tone, a call to action and a quality bar. The brief uprank writes has 18 fields: working title, intended reader, search intent, reader problem, primary question, unique angle, information gain requirement, section outline, required examples, required sources, internal link intents, external references, claims to verify, claims to avoid, visual requirement, cta, tone instructions, quality bar. 4 of them are optional, so the model can leave them out. The list fields are not optional but default to an empty list, so they can come back empty.
Does uprank write the content brief for me?
Yes. A model writes it as a step of every pipeline run, from the topic, the research and the brand profile, and a second model call reviews it against a checklist and sends it back up to 2 times. You cannot edit or approve it: the run goes from the brief straight to the draft, and the Brief tab on the article is read-only. brief_generate over MCP writes a brief on its own, but a later run does not use it.
Does uprank check the finished article against its brief?
Only in part. In a pipeline run the brief’s search intent is given to the model that judges the Search intent score, which is 10% of the overall score and has no gate of its own. The outline, required examples, required sources and claims to avoid are not compared with the draft, and the information-gain requirement is not compared with the element the draft declares. The 13 hard gates apply to the finished article whatever the brief said, so a person still reads the draft against the brief.

Keep reading

Create your workspace