Editing AI Drafts: A Subtractive Checklist for Content Leads

Most of us learned to edit assisted copy the wrong way round. A model hands back 1,400 words, the prose scans fine, so we do what we’d do with a freelancer’s draft: reword a clunky clause, swap a weak verb, break up a long sentence, move a paragraph. Forty minutes later the piece is smoother and no better. It still says nothing a reader couldn’t have guessed from the headline.

That’s the trap. Editing AI generated content is not a polishing job, because polish is the one thing the draft already has. What it lacks is subtraction and evidence. The sentences are grammatical and empty, and grammatical emptiness is invisible to line editing: you can rewrite “can help to significantly improve” into “improves” and you’ve fixed a hedge while leaving an unsupported claim standing.

So run a fixed pass instead. Three cuts, then one addition, in that order, every time. Deletion before insertion matters because cutting shrinks the surface area you have to fact-check, and because the holes left by deletion are exactly where the proof goes. I time this on our own drafts: a 1,400-word assisted draft takes 18 to 25 minutes through the subtractive pass, against 45 to 60 minutes of freeform editing that produces a worse piece. The pass is also teachable in an afternoon, which matters when you’re two people and a contractor.

Pass one: cut the throat-clearing

Open the draft and delete the first two paragraphs without reading them.

I mean that literally as a default, not a joke. Language models are trained to establish context before making a point, so assisted drafts almost always begin with 80 to 150 words of scene-setting that a reader who searched for the topic already knows. Check the specific phrases against your own back catalogue:

In today's fast-paced business environment,
As the digital landscape continues to evolve,
Whether you're a solo founder or an enterprise team,
Employee onboarding has never been more important.
In this article, we'll explore everything you need to know about
Before we dive in, let's define what we mean by

Delete to the first sentence that makes a claim. On the last 30 drafts I’ve run through this, the average cut was 112 words, and in 22 of them the real opening sentence was somewhere in paragraph three. You’ll often find the piece’s actual thesis buried at the end of the introduction, where the model finally arrived at the point. Promote it. That single move fixes more than any amount of sentence-level work.

Throat-clearing hides mid-piece too, at the top of every H2. Look for the pattern where a section opens by restating its own heading: under “Automating right-to-work checks”, the first sentence reads “Automating right-to-work checks can be a valuable step for HR teams.” Cut it. Go straight to the mechanism.

Pass two: cut the hedges

Now search, don’t read. Hedges are a finite vocabulary, which makes them the most automatable part of the job. Keep a find list and run it with Ctrl+F in Google Docs or a single regex in your editor:

can help to        may vary           it's worth noting
could potentially  might be           it is important to note
generally speaking often              many experts agree
in many cases      one of the most    some would argue
tends to           relatively         various
arguably           somewhat           a number of

Three things happen when you strip these. Sentences get shorter, so the draft drops 6 to 9 percent of its word count with no loss of meaning. Weak claims become obviously weak, because “automation can potentially help reduce admin time” stops being a comfortable sentence and becomes either “automation cuts admin time” (now prove it) or a deletion. And your tone stops sounding like a compliance document.

Two hedges are worth keeping, and only two: where the uncertainty is genuine and load-bearing (“HMRC has not confirmed whether the exemption applies to umbrella arrangements”), and where you’re describing a range you’ve measured. Everything else goes. If you use Grammarly or ProWritingAid, note that neither flags most of this: they’re tuned to clarity and grammar, not to epistemic padding. The find list beats the tool.

Pass three: cut unsupported claims

This is the pass that decides whether the piece is worth publishing, and it’s the one people skip because it’s the only slow part.

Go through the draft and highlight every factual assertion. Not opinions, not instructions: assertions about the world. Percentages, timings, causal claims, “studies show”, named tools and their features, prices, legal requirements. Then sort each one into three buckets.

The three-minute rule does the heavy lifting. Assisted drafts produce citation-shaped text: plausible percentages attached to plausible organisations, in the right register, with no underlying source. I have watched a draft attribute a retention statistic to Gallup that Gallup has never published, in a sentence so well-formed that two people signed it off. Set a timer, search for the primary source, and if it isn’t there in three minutes, cut the sentence rather than hunting for a replacement stat that says the same thing. Hunting for a stat to fit a claim you didn’t make is how you end up laundering an invention.

Expect this pass to remove 8 to 20 percent of the draft. On a recent 1,600-word piece about payroll integrations, 11 of 14 highlighted claims failed the test, and the draft came out at 1,050 words with four holes in it. Those holes are the point.

Pass four: add proof

You now have a short, blunt, honest, slightly gappy piece. Fill the gaps with things only your organisation can say.

Proof has a hierarchy, and it’s worth being strict about the order you reach for it. First, your own numbers: a client result, a before-and-after from your own operations, a figure from your GA4 or Ahrefs data with the date attached. Second, a named practitioner quote you can obtain in a Slack message, which for most in-house teams means a 20-minute conversation with someone in sales or support. Third, a primary source you’ve actually opened: a vendor changelog, ONS series, official guidance page, a filing. Fourth, and only fourth, a screenshot-free description of real tool output that a reader can reproduce.

Here’s the whole pass on one section. The draft:

In today's competitive hiring market, employee onboarding has become
increasingly important for organisations of all sizes. It's worth noting
that a well-structured onboarding process can potentially have a
significant impact on retention, and many experts agree that first
impressions matter. When it comes to onboarding software, there are a
number of factors to consider. Generally speaking, most companies find
that automating repetitive administrative tasks can help to free up HR
teams to focus on more strategic work. Studies show that 69% of employees
are more likely to stay for three years after a good onboarding
experience. It is important to remember that every organisation is
different, so your mileage may vary.

That’s 108 words. Pass one removes the first sentence. Pass two removes seven hedges. Pass three kills the 69% figure, which circulates constantly and traces back to nothing you can open. What survives is roughly 20 words of actual content, so pass four has to do real work:

Onboarding software pays for itself in one place: the admin hours it
removes from week one. Our own HR ops log gives the shape of it. Before
we consolidated contract signature, right-to-work checks and payroll
setup into one workflow in March 2026, each new hire took 3 hours 20
minutes of coordinator time spread across six days. After, it takes 45
minutes in a single sitting, and the coordinator no longer chases
documents by email. Retention is the harder claim, and we can't make it
from six months of data, so we won't.

Ninety-three words, all of them load-bearing, and the last sentence is more persuasive than the statistic it replaced. Refusing to claim something is proof of a kind. Readers are now extremely good at spotting a draft that claims everything and evidences nothing.

Running it with a team of three

The pass survives contact with a small team because it’s checkable. Someone can hand you an edited draft and you can verify the work in two minutes: did the throat-clearing go, does the hedge find-list come back clean, is every remaining number attached to a source you can click.

Track one number, the deletion ratio: words cut in passes one to three, divided by the original draft length. Healthy assisted drafts on a good brief lose 25 to 35 percent. Above 50 percent and the brief is the problem, not the editor, because the model had nothing specific to work with and filled the space with connective tissue. Below 15 percent and someone skipped pass three. We log it in a column in the Airtable content tracker next to the draft link, and it has been a better diagnostic of brief quality than anything else we’ve measured. Where the brief itself is the thing that needs fixing, that’s upstream work, and the drafting, brand voice and editing workflows pillar covers how the brief, voice guide and this editing pass fit together as one system rather than three habits.

One caution about tool choice: don’t ask the model to do the subtractive pass for you. I’ve tried it with Claude and with GPT-5.1 on identical drafts and both will happily cut hedges when instructed, which is the cheap 20 percent. Neither reliably deletes its own unsupported claims, because it has no way to distinguish a fabricated statistic from a real one, and both tend to replace a cut claim with a fresh one of the same shape. Automate the find list. Keep the judgement.

Next Monday, take the oldest assisted draft sitting in your queue, run the three cuts, and write the deletion ratio at the top of the document. Then go and read the brief that produced it.