What the red pen changes and what the naturalness score means
Updated
The red pen is an editing pass that runs after each draft. It checks 90 stock phrases in English and Turkish plus rhythm, sentence length, em dashes and repeated openers, edits what it finds, and shows a 0-100 naturalness score before and after: an editing score, not an AI detector.
Steps
- 01Open “Red pen” in the composer and pick “Off”, “Light” or “Strong” for this draft.
- 02To change the default, go to “Settings”, “Writing defaults”, “Red pen”. “Format default” keeps each format's own level.
- 03After the draft, read “Red pen notes”: the score before (struck through) and after, what was crossed out, and “Still there:” for anything left.
Off, Light and Strong
“Light” fixes the flagged phrases and any sentence that reads like a template, and leaves the rest as written. “Strong” rewrites every sentence in plain words and drops the general opening line and the closing summary. “Off” skips the pass. Both levels keep facts, names, numbers and bracketed placeholders, and aim to keep the length within about 15%.
Most formats start on Light, including Essay, Blog post, Anything and Rewrite. Fiction and Lyrics & poems start on Off; Product description and Ad copy have no red pen. Rewrite's Humanize always runs Strong, while Fix spelling & grammar, “Book outline”, “LinkedIn headlines”, “Resume bullet points” and Memoir's “Interview questions” skip the pass.
What it looks for
The phrase list has 54 English and 36 Turkish patterns, such as “a testament to” and “let that sink in”. Strong ones count on first use; mild ones only when repeated. Beyond phrases, the pass flags a flat rhythm (six or more sentences of similar length), an average sentence over 28 words, three or more em dashes at more than one per 120 words, the same first word on over a quarter of sentences, strings of rhetorical questions or exclamation marks, and piles of Turkish -mektedir/-maktadır endings. These findings go to the model as its fix list.
How the score works
Every text starts at 100 and loses points for each finding: 3, 7 or 12 per stock phrase depending on how strong a signal it is (at most 45 for phrases), 12 for flat rhythm, 8 for em dashes and 4 to 10 for the others. 85 and up shows as “Sounds human”, 65-84 as “Mostly natural” and below 65 as “Sounds generated”.
This measures editing habits that tire a reader. A high score says nothing about who wrote a text or what an AI detector will report.
What it doesn't do
- It makes no promise about any AI detector.
- Its phrase list is English and Turkish; in other languages only the rhythm, length, dash, opener and punctuation checks apply.
- If the pass is cut short, the draft from before it is kept with a warning.
Questions
Can I check a text without writing a draft?
Yes. The AI cliché checker page runs the same analysis in your browser and sends nothing to the server.