Claude style analyzer in depth

This Claude style analyzer produces a written assessment of a text: its register, the distribution of its sentence lengths, the constructions it repeats, and the points at which its voice changes. The output is prose. There is no score, no index and no dashboard.

The distinction is worth making explicitly, because the competing pages on this term are quantitative and a reader arrives expecting numbers. A formula returns the same value for the same input forever, which makes it comparable and makes it blind to meaning. An assessment reads the text, which makes it useful about things a count cannot see, and slightly different in wording each time you run it.

The analysis is generated for the text in front of it rather than filled into a fixed template, so emphasis follows what is actually notable. These are the recurring axes.

Register

Position on the formal to conversational range, and whether that position is held across the whole document.

Sentence length distribution

Whether lengths vary or converge. Convergence is the most legible signature of assisted drafting.

Constructional repetition

Recurring openers, connectives and clause shapes, counted by a reader's ear rather than by frequency.

Intra-document drift

Sections whose voice diverges from the rest, typically the seam between separate drafting sessions.

Audience calibration

Whether vocabulary and assumed knowledge match the reader the text addresses.

The route is server-backed, so each run consumes one of five in a rolling 24-hour window.

Step 1

Submit the text, between 1 and 12,000 characters. Submit the draft itself rather than an outline of it.

Step 2

Constrain the analysis with an instruction of up to 300 characters where you have a specific question, then run it.

Step 3

Act on one finding, revise, and only then spend a second run. Sequential passes on an unchanged draft add nothing.

No readability index is produced. Flesch-Kincaid, Gunning-fog, SMOG, Coleman-Liau and the Automated Readability Index are all absent, and where a grade level is a formal requirement this route does not satisfy it.

There is no lexical density figure, no parts-of-speech distribution, no word-frequency table or cloud, no n-gram listing and no passive-voice tagging in the text. Grammar and spelling are not checked. Nothing here constitutes a determination of authorship: style features describe a text, they do not identify a writer, and no page on this site will claim that they do.

Repeatability has real value. When you want the same input to produce the same finding every time, an assessment is the wrong instrument and a pattern match is the right one.

The watermark scanner on this site matches 17 named writing patterns and highlights each occurrence in place with the reason it was flagged: the scene-setting opener, the empty hedge, the formal connector, the overused verb, the stock metaphor, the hype adjective, the negation pivot and the rest. It executes locally, transmits nothing and carries no allowance, so it can be run on every revision. The two are complementary rather than competing: the scanner catches what is enumerable, and this analyzer addresses what is not.

Long documents are assembled across prompts, and each prompt carries its own framing. The model matches the framing it receives, so register shifts between sections in proportion to how differently each section was requested. The author, having read the text many times during assembly, is the person least able to notice.

Within a single response the tendency is the opposite, toward uniformity. Sentence lengths cluster, paragraph shapes repeat, connective vocabulary narrows. That regularity is what readers identify as machine-written, and it is difficult to self-diagnose because each individual sentence is unobjectionable.

Editors assessing whether a draft matches a publication's voice before it enters review, where the judgement would otherwise wait on a human reader's availability. Content leads auditing a set of pages for consistency across contributors and across months.

Analysts and consultants checking that a report holds one register from summary to appendix, since a document that changes voice halfway reads as assembled rather than authored. Students apply it to their own work under whatever policy their institution sets for AI assistance.