Claude tone analyzer without labels

The return value is a written account of how a passage is likely to land: which lines establish the register, where the register diverges from the situation the text is answering, and what a construction communicates independently of its propositional content. It is addressed to a person about to revise the text, not to a dashboard.

The route is server-side. Input is transmitted, accepted between 1 and 12,000 characters, and accompanied by an optional 300 character instruction. The allowance is 5 runs per rolling 24-hour window without an account. The analysis pass runs at temperature 0.2 against 0.55 on the rewriting routes, because a report that shifts between identical runs is not a report.

The established tools competing for this term are classifiers. One advertises up to 28 emotional tones, naming joy, anger, surprise, fear, admiration, curiosity, approval, disapproval and love, assigns them sentence by sentence, and prints a confidence figure with a published reading rule: above 70 percent is a strong signal, below it is subtle or mixed. Another reports a single dominant tone from a list of formal, friendly, persuasive and authoritative.

Classification is cheap to read, comparable between documents and trivial to chart. Those are real advantages and a page that pretends otherwise is selling. Its ceiling is equally real: a sentence labelled neutral at 68 percent does not tell you that the line reads as a reprimand because it opens by restating what the recipient failed to do. The label is accurate and useless at the same time.

This route produces no taxonomy. That forfeits comparability and buys specificity: the passage is quoted back, the mechanism is named, and the reader knows what to change. Neither model dominates. The choice is which failure you can afford.

Submit the complete text. Opening and closing lines carry disproportionate tonal weight and are, in most drafts, the passages actually responsible for the impression.

Step 1

Paste the full passage, 1 to 12,000 characters. The meter reports characters and words as you type.

Step 2

State the reader and their disposition in the 300 character instruction field, then execute. Tone is relational, so the same sentence has two different readings depending on who opens it.

Step 3

Revise the passages the report identifies. Remaining runs are printed on the widget, and a refused run names the reset time.

None of these is a grammatical fault, which is exactly why correctness tools pass the text through unflagged. They are reported consistently by guides written about Claude's default voice.

Post-assertion qualifier

A claim followed by a softener. "However, it's worth noting" is the canonical form. The proposition survives intact while the writer's commitment to it is withdrawn, and readers respond to the commitment.

Phatic opener ahead of content

"I hope this email finds you well" and equivalents. The cost is positional: the reader's attention is highest at the first line and it has been spent on nothing.

Symmetrical treatment by default

Acknowledging uncertainty and presenting multiple perspectives is the correct behaviour for analysis and a failure mode in correspondence that requires a decision, where symmetry reads as abstention.

Unrequested context

Background supplied beyond the scope of the request, typically placed before the request itself. It displaces the operative sentence rather than contradicting it.

Terminal offer of assistance

A closing line volunteering further work. Unremarkable in isolation, and a recognisable signature once it appears at the end of every message in a thread.

No tone label, because no tone vocabulary exists in this codebase to draw one from. No per-sentence assignment, because the route performs no segmentation and therefore has no sentence to attach a value to. No sentiment figure, no formality index, no confidence percentage: nothing in this path calculates a number, so publishing one would be fabrication.

Output is capped at 1,400 completion tokens, which matters at the top of the input range. A 12,000 character submission returns a report on the passages that carry the register rather than exhaustive coverage.

No browser extension, no mail-client integration, no API, no batch path, no file upload. One competitor in this result set advertises unlimited free analysis; this route is metered at five runs, so unlimited would be false here. A near-name competitor at a similar domain claims that text never leaves the device; that is also false of this route, which transmits. Stating the meter and the transmission is the only description of this page that survives inspection.

A model optimised for broad acceptability settles at the register that minimises objection across the widest set of readers: regular sentence lengths, hedged assertions, a positive terminal clause irrespective of subject. It is the register of institutional announcement, and it is emitted for a launch and for an outage without adjustment.

The resulting text reads as insincere rather than impolite, which is a harder failure to diagnose. Tone has stopped tracking content, and a reader registers the divergence between the news and the voice delivering it well before they can articulate the cause. That gap is what the report is built to name.

No. The system instruction governing this route forbids claiming certainty about authorship, plagiarism or which AI model produced a text, and requires observable signals and limitations to be described instead. That constraint is upstream of the page copy, not an appended disclaimer.

It is also correct on the evidence. Institutional human writing hedges and opens with pleasantries as house style, and edited model output frequently does neither. We make no prediction about any third-party detector's verdict and hold no affiliation with Anthropic.

The query is ambiguous in a way worth knowing. Searching the exact phrase returns a guitar tone fingerprinting repository, an audio analysis server, a music analysis skill and the Wikipedia entry for a French musicologist, alongside the writing tools.

Narrowing toward Claude writing pushes the set further off target: it fills with AI detectors, which answer a question about origin rather than about register. Almost nothing on the surface of this term is a tone tool for Claude prose specifically. This route reads written text only, has no audio path, and returns no verdict about who wrote anything.