The read-only word highlighter: seeing a draft before you touch it

Read-only is a feature, not a limitation

When you are judging a draft rather than writing it, the text has to hold still. An editable surface invites a small correction, then another, and within a minute you are no longer reading the piece you were asked to assess. Every count you took a moment ago referred to different words.

A read-only preview makes the reading stable. Marks are rendered from the pasted text and nothing can change underneath the cursor, so the density figure at the top of the page still describes the paragraph you are looking at. That sounds like a small thing until you have tried to compare two drafts in an editor and lost track of which version produced which number.

It is also the honest way to review a model's output. The temptation with generated text is to improve it immediately, which hides the evidence of how much improvement it needed. Marking without touching shows you the size of the editing job before you start spending the afternoon on it.

For agency work there is a third reason. A read-only pass is reproducible. Two people can paste the same text and see the same marks in the same places, which makes the review a shared artefact rather than an opinion about a document that keeps changing.

The five pattern sets and when to use each

The tells set is the general-purpose one: vocabulary a model overuses, hedges, announcement connectives and unproven absolutes. Start here on any draft whose provenance you do not know. It is the set that best predicts how much line editing a piece will need.

Buzzwords is the marketing set. Set phrases, aspirational verbs, the vocabulary of transformation. It is less about detecting a machine and more about catching copy that says nothing a buyer could act on, which is a commercial problem before it is an authenticity problem.

Claims is the set with teeth. Vague attribution, unsourced statistics, placeholder citations and the shapes that precede a number nobody checked. Run it on anything factual: a report, a policy, a product description that makes a performance promise.

Overused words and filler is for tightening prose that already reads well. It catches the verbs and adjectives that survive every rewrite because they feel professional, and the filler phrases that pad a sentence to the length a template expects.

• Tells: unknown provenance, general line editing.
• Buzzwords: marketing copy, landing pages, sales sequences.
• Claims: anything factual, regulated or contractual.
• Overused words: a fourth pass on prose that is already clean.
• All four: a final check before publishing, with the marks read as one picture.

Reading density rather than counting hits

A single flagged word is noise. Twenty in eight hundred words is a signature. The panel reports flags per thousand words precisely so that length stops distorting the judgement, and so a long document does not look worse simply because it is long.

Under three per thousand, treat the marks as suggestions. Between three and eight, the reader is already feeling a rhythm they cannot name, and the piece will read as competent and forgettable. Above eight, individual substitutions will not help. The paragraphs need rebuilding around specifics, and the vocabulary will fall out of that work on its own.

The most useful comparison is against your own previous draft. Paste version one, note the density, edit, paste version two. If the figure fell by half and the prose is genuinely better, the edit did what you hoped. If the figure fell and the piece still feels hollow, you replaced words and left the emptiness in place.

Two cautions. Density is not a grade, and a technical document with legitimate repetition will score higher than a magazine feature that says less. And a clean score is not an accuracy check — a hundred percent factual error rate is perfectly compatible with zero flagged vocabulary.

Working through a pass in the right order

Read the piece once with all sets on, purely to see the shape of the problem. Do not fix anything. You are answering one question: is this a word-level draft or a paragraph-level one? That answer decides how you spend the next hour.

Then turn sets off until only the risky one is visible. Claims, if the piece makes factual assertions. Work every mark: source it, cut it, or rewrite it as a smaller claim. Turn that set off before moving on, so the next family is not competing for attention.

Then the tells, then the vocabulary. The order is risk first, style last, for a practical reason: attention is a finite resource and style work is the most seductive form of procrastination available to a writer.

Finish by reading the text with no marks at all. This is the step everyone skips, and it is the only one that answers the question the reader will answer: does this read like something a person meant?

Pasting in chunk sizes that fit a judgement

Density normalises for length, which makes it tempting to paste a whole report and read one number. Resist that. The average of five thousand words hides the section that is three times denser than the rest, and the section is where the work is.

Paste a section at a time: a heading and its paragraphs, or a landing page, or a product description. The figure is then about something you can hold in your head, and the edit is something you can finish in the same sitting.

Keep headings out of the pasted text where you can. They are usually shorter and denser in flagged terms than body copy, so including them tilts the density upward, and a document with many subheadings will look worse than the same words in fewer sections.

The other rule is to compare like with like. Two versions of the same section tell you whether an edit worked. A section against a page tells you nothing, because the difference between them is mostly the subject.

House rules for a client, and a baseline

Most client style guides reduce to somewhere between six and ten words: no exclamation marks, never say cheap, always capitalise the product family name, avoid the word solution. Those are ideal for a literal phrase list, and once they are written down the review stops depending on whether the freelancer happened to read the brand book.

The second thing worth recording is a density baseline. Some brands write in a style that scores high on these lists by design, and a reviewer who does not know that will spend an afternoon removing vocabulary the client paid for.

Take the baseline from three pieces the client has already approved, average the figures, and put the number in the brief next to the word count. A new writer then knows what normal looks like for this client rather than for writing in general.

Review the baseline when the style changes, and note the date. A number with no date becomes folklore within a year, and folklore is what a proper style guide exists to replace.

Where the highlighter sits in a real review

The order that works is cheapest check first. The console runs structure, claims and formatting in a second. The highlighter reads the prose. The editor is where you spend the afternoon. Running them in that sequence means the expensive pass never happens on a draft that will be rejected for a missing section.

It also makes a decent gate before human review. Sending a draft at a density of eleven to a specialist reviewer is asking them to do vocabulary editing at a professional rate; sending the same draft after a highlighter pass means the reviewer's time goes to the parts that need their judgement.

Keep the density figure with the finished piece, alongside the automated report. When a client asks six months later whether the copy was edited, a number taken at publication is a fact, and a recollection is not.

A number worth putting in the brief

When you commission copy, state the density you expect and how it will be measured. A freelancer who knows the figure is being checked writes to the standard, and a disagreement about whether a draft sounds generic becomes a comparison of two numbers rather than a matter of taste.

The number should be a ceiling rather than a target. Copy engineered to score zero on a vocabulary list reads like copy engineered to score zero on a vocabulary list, which is its own kind of generic, and the reader notices the absence of any risk taking long before they notice the absence of any particular word.

What a flagged word does not prove

No list on this page proves a machine wrote the text, and the tool will never claim one did. Human writers use every word here, often badly, and the lists were built from prose written by people as much as by models. Detectors that promise otherwise fail predictably on careful writers, formal registers, non-native speakers and anyone writing in a second language.

The list is an editing instrument. Its job is to point at the places where a sentence could be more specific, and its value is measured by how many of those places you can improve, not by how confidently it accuses the text.

It also cannot see the failures that matter most: a number with no source behind it, a citation to a document that does not exist, a quotation nobody said. Those need a source comparison and a claim ledger, which is a different kind of tool doing a different kind of work.

Use the highlighter to decide whether a draft needs an afternoon or a rewrite. Use the console and, where it counts, a named reviewer to decide whether it is true.