Editing generated content is six passes, in a fixed order.

An editing pass is a decision that applies to every sentence, made once so it does not have to be remade each time. Facts first, then structure, then wording.

A generated draft has a predictable profile: fluent, evenly weighted, confident where it should hedge, unsourced where it should cite, and longer than the point requires. Each of those is a pass.

Editing in the wrong order is what makes the task feel endless. Rewriting a sentence that will be deleted two passes later is the most common way to spend an afternoon on a page that gets no better.

The six passes here take about an hour on a 1,500-word draft. The twenty-minute version is at the end for when that is not available.

Every factual sentence gets one of three marks: sourced with a link or reference, verified and true, or deleted. There is no fourth category, and probable is not one of the others.

The sentences that need the closest attention are the confident ones with a number or a date, because models produce these in the same register as generalities and they are the ones a reader will repeat to someone else.

Check dates as well as facts. A figure from four years ago is not wrong, it is outdated, and an outdated figure presented as current is a factual error once the world has moved.

Anything you cannot verify in the time available goes. Cutting a claim costs nothing; publishing a wrong one costs the page's credibility and, if it is repeated elsewhere, more than that.

Each heading should be answerable in one sentence, and the first sentence under it should be that answer. If a heading cannot be answered, it is a subject the draft did not have material for, and it should be cut.

Then run the deletion test on every paragraph: remove it and ask whether a reader lost a fact. Paragraphs whose removal changes nothing are the padding that generation adds to fill a word count.

Check the order. A page that defines a term three sections after using it, or explains a constraint at the end, is ordered by the template rather than by the reader's needs.

Delete the conclusion if it restates the introduction. A conclusion earns its place by adding a decision, a limit or the next step, not by summarising what the reader just read.

This is the mechanical pass, and four categories cover most of it: hedges, empty intensifiers, restated conclusions and wordy phrasing.

Hedges are the qualifiers a model uses when it is not sure but does not want to say so: it is important to note, arguably, generally speaking. Each one either disappears or is replaced by the specific condition that makes the claim true.

Intensifiers stand in for evidence: very, highly, significantly, extremely. If something is significant, the number in pass one shows it. If there is no number, the word is doing work nothing supports.

Wordy phrasing has a plain equivalent, and the plain version is almost always better: in order to becomes to, due to the fact that becomes because, at this point in time becomes now.

Now the page reads correctly and still sounds like a category. Four checks make it sound like a person: the opening, the constructions, the emphasis and the punctuation.

The first sentence should carry a fact or a claim, not announce the article. Delete any in this article, in today's landscape or here is what you need to know opening and start with the second sentence.

Count the contrast formulas and promise formulas. One of each is style; three is a template. Choose the one that matters and rewrite the others as plain statements.

Then find the sentence in each section a reader must not miss, and make it short and unqualified. Everything around it can hedge; that sentence cannot, or the section has no point.

Format is what makes the page legible in a result, a reader and a screen reader. It is also the pass most often skipped, because the page looks finished by now.

Title and meta description: does the title state what the page delivers and the description say why it is different? Neither should repeat the other.

Headings: in order, no level skipped, no heading used only for visual size. Lists: used only where the items are parallel. Links: descriptive text, not click here.

Images, tables and code: alt text that conveys the content, tables with headers, code blocks that can be copied. Check the page at a narrow width, because a table that only works at desktop is a broken table for most readers.

Last, and the pass that makes the other five checkable: who is responsible for this page being right, and when was it last reviewed?

A named author with a real profile, an update date, and sources where the claims need them. If the page affects money, health or safety, a note on how it was reviewed and by whom.

A short process note is enough: drafted with assistance, fact-checked by a named editor, sources listed. It answers the reader's actual question, which is whether they can trust the page.

Then record the edit itself, in one line: what changed and why. A page without a history gets re-edited from scratch by the next person.

When the full pass is not possible, do three things in this order: delete every unverifiable claim, delete the padding, and rewrite the first paragraph so it states the point.

That sequence addresses the failures that cost the most: a wrong fact, a page that says nothing, and a reader who leaves before the substance starts.

Skip the voice pass entirely on that schedule. A page that is dull but accurate is worth publishing; a smooth page with a wrong figure is not.

Leave a note of what you skipped, so the next edit knows the page has had a partial pass rather than a full one.

Do not reword a claim you cannot verify. Rewording makes an unsupported statement harder to spot without making it true, which is worse than leaving it and flagging it.

Do not smooth away the specifics. An editor's instinct is to turn a measured result into a general trend because it reads better, and that instinct removes the only original material on the page.

Do not polish the prose until the structure is right. A well-written page in the wrong order is still the wrong page, and the time spent on the sentences is wasted when the sections move.

Do not edit towards a detector score. Editing towards accuracy and specificity produces a page that is defensible; editing towards a percentage produces one that is merely less suspicious.

A checklist with forty items gets ignored. Six passes with three to five checks each is what people actually complete, and completion matters more than coverage.

Give the checklist an owner, who is accountable for outcomes rather than for enforcing the process. The measure is whether errors reach readers, not whether the boxes were ticked.

Review it quarterly against what actually went wrong. Remove the checks that never catch anything and add the one that would have caught last month's mistake.

Keep the record of what the passes found. After three months it tells you which failure your drafts have, which is the only information that improves the brief.