How to Edit AI Content: The Feedback Loop Nobody Runs

If your AI writing constantly elevates the user experience, it isn't just this, it's that, and you're inviting your readers to dive in, then your AI copy sucks.

There are three tells in that sentence and I put them there on purpose. A model uses them because millions of marketing pages used them first.

Editing AI content is a different job from editing a person's work. A person makes a mistake once and you can explain it to them. A model makes the same few mistakes in every draft, forever, until something stops it. So the job isn't proofreading. It's building the thing that stops it.

What It Means to Edit AI Content

To edit AI content properly you run three passes, and most people only run one of them.

The first asks whether it's true, and it's the one that can hurt you. A model states a wrong fact in the same confident tone it uses for a right one, so nothing in the writing tells you which is which.

The second asks whether it sounds like you rather than like everyone else. The third is a proofread, which catches typos, and models rarely make typos. Most teams run that third pass and skip the first two, which is why so much reviewed AI content still reads like a machine wrote it: it was checked for the one thing the machine is reliably good at.

The Tells Worth Hunting

Six patterns cause most of the problem, and you can find all of them with a text search rather than a judgement call.

• Verbs that mean nothing. Elevate, leverage, streamline, foster, unlock, harness. None of them describe an action. Replace each with what actually happens.
• The false distinction. "It isn't just a tool, it's a partner." It looks like a clarification and adds nothing. Make the claim once.
• Lists of three. Fast, simple and secure. Models reach for three because most writing does. Cut one item.
• Sentences that are all the same length. This is the biggest tell and the least discussed. People write a six-word sentence, then a forty-word one. Models stay at about eighteen.
• The sentence that repeats the paragraph. It arrives at the end and restates what you just read. Delete it.
• Hedging words. Typically, generally, often. They soften a claim the writer wasn't sure about. Either make the claim or drop it.

None of these need judgement. They need a list, and a list can be given to something that checks every time.

A person makes a mistake once and you can explain it to them. A model makes the same few mistakes in every draft, forever, until something stops it.

Three Files Run Against Every Piece

I use Claude for this. The control doesn't come from the prompt. It comes from three files, and every piece is written against them, fixed against them and scored against them.

A prompt is something you ask for once. A file is something that applies whether or not anyone remembered it on a Friday afternoon, and most of the quality gap in AI content sits in the space between those two things.

Why a File Beats a Prompt

Prompting is where most advice about AI writing stops, and it runs out within a fortnight for three fairly practical reasons.

Prompts get retyped. What you asked for last week is gone unless someone pastes it again, and what gets pasted is a shorter version with the clause nobody remembers mattering taken out of it, whereas a file gets read in full every time, including that clause.

Prompts can't be compared. When the writing gets worse you want to know what changed, and with a prompt the honest answer is that nobody is sure. With a file, it's a line and a date.

Prompts are too short. Fifty-eight banned words, fourteen structural patterns and a contraction policy that changes by page type is not something anyone pastes into a chat box every morning, so in practice it never gets applied at all.

The Standard: What Good Looks Like

The first file is the writing standard. It covers voice, banned words, the patterns that aren't allowed, and how a claim has to be sourced. It runs to just under eleven thousand words and it's dull, which is right for a rulebook.

What matters is that each rule is written so you can tell whether it was followed. "Sound conversational" isn't a rule, because two people will read it differently. "No em dashes, use a comma or a full stop instead" is a rule.

The Protocol: How a Fix Runs

The second file sets out how to fix a page: structural read, fact-check, voice pass, the edits, then a check against the live page rather than the file on your machine.

That last step matters more than it sounds, because the common failure isn't a bad fix. It's a fix that didn't save properly and nobody looked again in the same pass.

The Audit: What Shipped

The third file scores the published page and changes nothing. It gives a number, a list of problems, and a threshold below which the page doesn't go out.

Keeping the audit separate from the editing matters. If the same pass can find a problem and fix it, it stops being independent.

What That Control Gives You

The model doesn't get any better at writing. What you get is the same result every time, and a way to settle an argument about a draft without either person appealing to taste.

Once the rules are written down, "I don't like the opening" becomes a question you can answer by checking it against the rule on openings. Either someone broke the rule or the rule is wrong, and you can fix both.

The other gain is speed on the dull work. Nobody has to remember fifty-eight banned words, because the check remembers them, runs in under a second, and hands back line numbers.

Here is one that earns its place every week. The rule says no bracketed asides in body prose, on the grounds that a writer who puts a thought in brackets did not trust the sentence around it. That is a preference, and reasonable people disagree with it. But because it is written down and checked, nobody has to have the argument twice, and the piece you are reading now contains none.

The Feedback Loop

This is the part almost nobody runs. Every correction you make by hand is a rule you haven't written down yet. Write it down and you won't make that correction again.

Most teams edit the draft, publish it, and then start the next one from exactly the same place, because nothing about the system changed in between.

1. Edit the draft as usualFix what's wrong. Don't think about the system yet. Ordinary editing
2. Notice what you've fixed twiceIf you've corrected the same thing in two pieces, it's a pattern.The signal
3. Decide whether a machine could catch itA banned word, yes. A weak argument, no. The split
4. Write it into the standardOne sentence, worded so someone else can tell whether it was followed.Now it's a rule
5. Add it to the checkerIf it's mechanical. A search, a count, a limit. It runs on everything after that.Now it's a gate
6. See what it catchesA rule that never fires is either working or pointless. One that fires constantly is worded badly.The loop closes

Here is a real one from this site. Early drafts kept opening paragraphs with a short fragment, which reads like a report rather than a person, and I corrected it by hand in three pieces before noticing the pattern. It went into the standard as a rule about openers, and into the checker as a count of how many paragraphs in a row start the same way. It has not come back since.

After a couple of months the drafts arrive in better shape. The model hasn't changed at all, but the instructions it works from now contain everything you have objected to before, so what used to be corrections have quietly become the starting position.

And the honest part. This removes the mechanical faults and does nothing for the interesting ones. A model that has stopped saying leverage still has nothing to say.

The Passes, In Order

Order matters. Fix the voice before the facts and you'll polish sentences you're about to delete.

Two of the six are mechanical and the other four are judgement, which is worth sitting with given how much of the marketing around AI editing implies the reverse.

One Paragraph, Before and After

Here's a specimen, and I wrote the first version to contain all six tells, which took no effort at all, because asking for a paragraph about a product with no other instruction gets you most of the way there anyway.

Before: "Our platform is designed to elevate your workflow and streamline collaboration across teams. It isn't just a project tool, it's a partner in productivity. Typically, organisations see faster delivery, better visibility and improved morale. By leveraging automation, teams can focus on what matters most."

After: "The platform does one thing. Everybody on a project sees the same list, in the same order, updated at the same time. Teams that move to it stop holding the Monday meeting where everyone says what they're working on. That meeting was ninety minutes."

Forty-two words became fifty-one, so this isn't about cutting length. Five things changed.

The last row is the important one. There's nothing in the original a reader could argue with or remember, and that isn't because a machine wrote it. It's because nobody gave the machine anything only they knew. The edit put a fact in.

Fact-Checking Is the Pass You Can't Automate

A model will give you a precise number with nothing behind it, in the same confident tone it uses for the real ones, and no voice check will ever catch that.

The rule I work to is that every disputable claim gets checked the same day, against a primary source. Not from the model's memory and not from mine. Figures, dates, named organisations, anything attributed, and every comparative, because "the largest" is the riskiest phrase in commercial writing.

One thing to be careful of, though. Checking a number isn't the same as citing it, and being unable to link a source isn't a reason to cut it. If several independent write-ups agree on a figure, that figure is a fact and you can state it. A page with no numbers loses to a page with numbers.

Does Edited AI Content Still Rank?

Yes, and the question has a wrong assumption inside it. Google's spam policy is aimed at content made at scale to game rankings, and it says nothing about whether a machine helped.

What does go with poor performance is the thing the tells point at: writing that is uniform, vague, hedged and impossible to check. Pages like that did badly before anyone could generate them quickly. I've written about that in more detail, with the study figures.

So none of this is about avoiding a detector, and every pass in the table above would improve a piece written entirely by hand.

When to Stop Editing and Start Again

Some drafts can't be fixed, and noticing early saves a day, so the test is whether there's an argument in there. If there is one and the writing is poor, edit it. If the writing is fine and there's nothing underneath, no amount of editing will put an argument there, and you are about to spend three hours making an empty page sound better.

You'll know because you keep improving sentences and it still isn't about anything. Stop, write down what you think in six bullet points, and start again from those.

What None of This Fixes

A standard, a protocol and a check will give you clean, consistent copy that doesn't read like a machine, but they won't give you an opinion, and the opinion is the thing that makes a page worth reading in the first place.

Nothing in the system knows what you believe, which customer keeps asking the awkward question, or what your competitors get wrong. Someone has to put that in. The files protect the floor. They don't raise the ceiling.

So the useful version of this is fairly plain. Write the rules down, let a machine check the half that can be checked mechanically, and spend the time you save on the half that can't. The pages that come out the other side aren't good because of the system, they're good because the system stopped spending your attention on the word leverage.