Making AI a Writing Partner That Earns Its Place
See where an AI writing partner earns its place across research, drafting and editing, and which parts of the writing stay with you.
By David Jubé · Jun 25, 2026 · 14 min read
Prompt craft is the wrong lever: an AI writing partner earns its place tier by tier, and the tiers are research, drafting and editing.
Which parts of a draft need a person, and which parts only need volume? Research and first-pass structure need volume. Point of view, first-hand detail, and an example that could only belong to your business do not.
That split is easy to state and easy to lose, because it does not sit still. It moves between the brief and the publish button, and it reverses once, at the edit.
Key takeaways
- Split the work by tier, not by draft. A model earns a lot at research, some at drafting, and almost nothing at the edit.
- Ask a model for the disagreement in a field, not the summary. Consensus is what you write against, not what you publish.
- Hand over the paragraph that reads the same on anyone’s site. Keep the paragraph that could only be true of yours.
- The edit pass is subtraction: cut the hedge, cut the sentence that survives its own deletion, swap the generic example for a real one.
- Three schedule failures break the method: a brief with no angle, an edit skipped under deadline, and nobody reading the page before it goes out.
Division of Labor, Stated in One Paragraph
Working collaboration divides on one axis: a model drafts structure and first-pass volume from a brief that already carries an angle, and a person supplies the point of view, the first-hand detail, and the example a model has no way to invent.
That is the answer in one paragraph. The rest is the mechanism, tier by tier, plus what breaks when the two halves swap.
What makes it hard is not the principle. It is that a page is rarely one kind of writing.
The paragraph defining a term your reader half-knows is volume work. The paragraph explaining what you charge, and why you changed it last spring, is not. Handing both to the same process, in either direction, is where this fails.
The ratio is different at each tier of the work:
- Research, where a model does most of the lifting
- Drafting, where it does about half, and only the structural half
- Editing, where it does almost none
The tool question is largely settled in practice already. In the 2026 B2B content marketing research from CMI and MarketingProfs, a survey of 1,015 B2B marketers fielded between June and August 2025, 95% said their organization uses AI-powered applications, so the live question is no longer whether to use one but what to hand it.
Why AI prose reads the way it does is its own subject, covered in use AI to write without sounding like AI. This piece closes the cluster on the process instead: which tier of the work a model can hold, and what a person adds at each one.
Research Is Where an AI Writing Partner Earns the Most Time
Research is the tier where a model’s biggest weakness turns useful. Trained on the average of everything written about a subject, it hands back the consensus position in seconds, and you need that consensus before you can decide where to break with it.
Consensus is cheap to read, slow to assemble by hand, and worth the trade every time.
Ask for the Disagreement, Not the Summary
Summarizing ten sources tells you what they agree on, which is the least useful thing in the pile. The angle lives in what they say differently, so ask for that instead.
Four asks do most of the work at this tier:
- The contradiction: where do these sources disagree, and on what exactly.
- The unsourced repeat: which claim appears in all of them with no primary source underneath it.
- The objection: what would a skeptical reader push back on at each section of this outline.
- The unanswered question: what does a reader want to know after the third section that the outline never covers.
The second ask is the one that pays. Repeated-with-no-source is how folklore travels in this industry, and a model is fast at spotting the repetition even though it cannot judge the claim.
Pressure-Test the Outline Before You Write It
An outline is cheap to break and expensive to rebuild after drafting. Asking a model to argue against each section before any prose exists is the cheapest edit available to you.
What you feed it at that moment does more work than how you phrase the request, which is the point behind Anthropic’s framing of context engineering as the successor to prompt engineering: curate the material the model reasons over, rather than polishing the wording around it. OpenAI’s own docs make the same split mechanical, with a standing instructions layer that takes priority over whatever you type into a given request.
None of this asks the model to know anything true about your business, only to be fast and untroubled by having its work thrown out.
This tier assumes the brief already exists. A model asked to research a topic rather than a brief returns the most average version of that topic, which is why the brief with an angle comes first and reading the results comes before that. Those briefs come out of the content audit that produces your next ideas.
Drafting: Volume Is the Job, Judgment Is Not
Drafting is where the split gets blurry, because a single section usually contains both kinds of writing.
Hand over the structural parts. The definition a reader needs before the argument can start, the connective tissue between two sections, a first pass at something you have explained twenty times already: none of that carries evidence, and all of it takes time.
Keep the evidentiary parts.
Take a hypothetical two-person studio writing a page about lead times. The paragraph explaining what a lead time is can be drafted by a model, because that paragraph reads the same on every site in the category. The paragraph explaining why this studio quotes six weeks while the category quotes four cannot be, because the reason sits inside that studio’s own operations.
One page, two jobs, and only one of them is a model’s.
The Test That Decides Each Paragraph
The practical test at this tier is a question about the sentence in front of you, not about the article:
- Could this paragraph appear unchanged on a competitor’s site? Hand it over.
- Does it carry a number, a name, a price, a decision, or a preference? Write it yourself.
- Does it state the position the page is willing to defend? Write it first, before anything else gets drafted.
That last one matters more than the order suggests. A model given a position will hold it through a draft; a model asked to invent one defaults to whatever is already most common in the results. Anthropic’s own business guidance lands in the same place, recommending you treat the model like an intern on their first day and give it the full instruction rather than a topic.
Few people are handing over the whole job anyway. HubSpot’s State of AI report, a survey of more than a thousand marketing and advertising professionals, found only 4% use AI to write entire pieces of content, with outlines and first ideas the far more common use. That ratio is roughly what the tier split predicts, and the limits these tools hit on their own are the reason it holds.
Founder Insight The tier that gets cut under deadline is the edit, and the edit is the tier that puts a specific business into a general draft. If the week only has room for two of the three tiers, I would drop the drafting tier and write that part myself.
Editing Reverses the Split, and That Is the Point
Editing is where the ratio inverts. At research a model does most of the lifting; at the edit it does almost none, because the work is subtraction and judgment about one specific business.
The Three Cuts That Do the Work
This pass is not a polish. It removes three things, in this order:
- The hedge: any phrase that softens a claim until it commits to nothing, replaced with a number, a name, or a stated position.
- The sentence that survives its own deletion: if the paragraph loses nothing when the sentence goes, the sentence was filler.
- The generic example: any example that could describe a different business, swapped for one that could only describe yours.
The third cut is the slow one, and it is the one that exposes whether a paragraph was ever about anything.
Splitting the pass in two helps, and the Content Marketing Institute’s guidance on editing AI content sets it out that way: a strategic pass for argument and fit, then a line pass for the sentences. Animalz makes the harder version of the same argument, that AI drafts arrive carrying generic claims and confident errors that only editorial review catches. Neither is optional if the draft is going out under your name.
There is a reading-behavior reason the hedge cut comes first. Nielsen Norman Group’s eyetracking work on the F-shaped reading pattern found that the first words of a line take disproportionately more attention than the rest, so a paragraph that opens on a qualifier spends its most-read words saying nothing.
The One Editing Job Worth Handing Back
Once a draft exists, a model is good at one narrow editorial task: naming what is missing. Ask which claim in the draft has no example under it, and which question a reader would ask next that the draft never answers.
It will not supply either answer. Naming the gap is the whole contribution, and it is a real one, because a writer three hours into their own draft has stopped seeing the holes in it.
Proof a Model Cannot Manufacture
Three things stay with a person at every tier: a position the page is willing to defend, a detail from something that happened, and an example specific enough that a competitor could not run it unchanged.
That is not a modesty point about what models cannot do, and I do not expect a better model to change it. It is a description of what those particular words are made of.
A model trained on the average of a subject returns the average position, the average example, and the average level of specificity. Grow and Convert calls the output mirage content, articulate and empty at the same time, because the statistically average answer is what the machine is built to produce. Useful in a first pass, disqualifying in a finished page.
Three questions separate proof from padding:
- Who decided this, and what did they decide against
- When did it happen, and what did it cost
- Could a competitor put their own name on this sentence
The difference between a writing partner and a ghostwriter is accountability rather than output. A partner drafts what you can check line by line; a ghostwriter hands back a finished page you either trust or do not.
Why Readers Register the Difference
Readers are not neutral about which one they got. In the Reuters Institute’s six-country survey on generative AI and news, fielded in mid-2025, 12% said they were comfortable with news made entirely by AI against 62% for news made entirely by people, which is the audience any proof on your page has to satisfy.
Before anything publishes, run the two tests this cluster is built on. Does the page serve the person reading it ahead of the ranking it is chasing, and does it carry proof a reader and a quality rater can both see? Both are defined in full in what Google actually penalizes about AI content, and the second is the ground proving experience on the page covers article by article.
Where This Method Breaks, and What Catches It
Three failures account for nearly every AI-assisted page that reads like nobody was home. All three are schedule failures rather than judgment failures.
Three Points Where It Fails
- A brief with no angle: the drafting tier is forced to invent one, and a model asked to invent an angle returns whatever already sits in the results.
- An edit skipped under deadline: subtraction is the step that gets cut when time runs short, which is how a competent draft ships with no evidence in it.
- No second reader: the page goes out having been read once, by the person least able to hear it.
That third one is the least intuitive and the cheapest to fix.
A writer who has read their own opinion ten times stops hearing it as a claim and starts hearing it as background. A second reader hears it once, which is the only condition under which a hedge is audible.
The Check That Catches Each One
None of this is a tooling problem, which is why a new model rarely fixes it. It is a process that stopped checking its own output, and the repair is to put the tiers back in order and run them. Ahrefs makes that case from the writing side.
Semrush makes it from the optimization side, treating reader-facing and search-facing quality as one job rather than a trade. Where this method sits inside the full pipeline is covered in how to produce content that ranks and gets cited, and the shape a finished page should hit is in content that ranks and gets cited by AI.
Frequently Asked Questions
What should AI do at the research stage?
Summarize the consensus, then name where sources disagree and which claims repeat with no primary source underneath. That is the fastest useful output at this tier. It also pressure-tests an outline: ask what a skeptical reader would object to at each section before you write a word of prose.
How much of a first draft can a model reasonably write?
The structural half: definitions, connective tissue between sections, and explanations that would read the same on any site in your category. Anything carrying a number, a price, a decision, or a preference stays with you, because that is the half a reader cannot get anywhere else.
Which cuts does an edit pass on an AI draft need to make?
Three, in order: the hedge that softens a claim until it commits to nothing, the sentence the paragraph loses nothing by deleting, and the example that could describe any business in your category. The third is slowest and matters most, because it exposes paragraphs that were never about anything.
Do better prompts produce better articles?
Only up to a point. A prompt cannot supply a position, a first-hand detail, or a specific example, so past a reasonable standard of clarity the returns fall off fast. What moves quality is the brief behind the prompt: a stated angle, a named reader, and the claim the page is willing to defend.
Who should the byline credit on an AI-assisted page?
The person accountable for the claims, since a byline is an accountability signal rather than a typing credit. Whoever checked the facts, supplied the examples, and would defend the page in a room owns it. Disclosure rules are a separate question, and a legal one rather than an editorial one.
Where does a person stay irreplaceable in this process?
At three points: setting the angle before drafting, supplying first-hand detail a model has no access to, and running the subtraction pass at the edit. Those three are the reason a page reads as though somebody was accountable for it rather than merely present at its production.
Continue Reading
More From This Series
- What Google Actually Penalizes About AI Content
- The Evidence of Effort a Reader Can Actually See
- The AI Tells That Cost You Readers, Not Rankings
- Scaled Content Abuse Is About Quality, Not About AI
- The Content Audit That Produces Your Next Ideas
More from TDM Insights
- Use AI to Write Without Sounding Like AI
- How to Write a Content Brief (That Gets It Right)
- How to Read the SERP Before You Write a Word
- Writing for E-E-A-T: Prove Experience on the Page
- How to Produce Content That Ranks and Gets Cited
- Content That Ranks and Gets Cited by AI
Explore TDM Insights Topics