The check
What the Claude tone analyzer checks
This is a pre-send check. You have a draft that came out of Claude, it is grammatically fine, and you are not sure how the person on the other end will take it. The Claude tone analyzer reads it and reports back in writing: which lines set the register, where the register works against the message, and what a given phrase signals to a reader who did not see your prompt.
It runs on our AI server, so the text is transmitted when you press Analyze. The allowance is 5 runs per rolling 24 hours, the input box takes 1 to 12,000 characters, and there is a 300 character field for telling the analyzer who is reading. Nothing to install, nothing to sign up for.
Workflow
When in the workflow to run it
Five runs a day is enough for a working week if you spend them on the drafts that actually carry risk. These are the moments worth one.
- Before a reply to an unhappy customer
- The draft will be correct. The question is whether it reads as an apology or as a defence, and that turns on two or three lines.
- Before written feedback
- Run it with the recipient described in the instruction field. Encouragement and warning are separated by very little on the page.
- Before a decision goes out
- If the analyzer cannot find the recommendation, neither will your reader. That is the useful failure.
- After a heavy edit
- Editing half a draft and leaving the other half often produces two registers in one message. This is the check that surfaces the seam.
- Not on internal one-liners
- A three sentence Slack message does not need a run. Save the allowance.
How to
How to run the Claude tone analyzer in three steps
Paste the message end to end. Greeting and sign off do more tonal work than the body, so a paste that starts at paragraph two hides the thing you are looking for.
- Step 1
- Paste the full draft, 1 to 12,000 characters. The counter above the box tracks characters and words live.
- Step 2
- Fill the instruction field with the situation in 300 characters or fewer. Who is reading, what they already know, and what mood they are in when they open it.
- Step 3
- Press Analyze, then edit only the passages the report names. Copy the report if you want it beside the draft while you work.
Patterns
The Claude tone patterns to look for yourself
Some of this you can catch before spending a run. Writing guides describing Claude's default voice keep naming the same handful of constructions.
- Hedge on the tail
- The assertion arrives, then a qualifier softens it. One guide quotes "however, it's worth noting" as the standard shape. Delete the qualifier and check whether anything was lost.
- Warm-up before the point
- "I hope this email finds you well" and its relatives. Cut the first sentence and see whether the message still opens correctly. It usually opens better.
- Perspectives instead of a call
- Acknowledging uncertainty and laying out both sides. Right for a memo, wrong for an approval request.
- Volunteered background
- Context you did not ask for, placed ahead of the ask. Move the ask up and see how much of the background survives.
- The offer at the end
- A closing line offering further help. Fine once in a thread, filler by the third reply.
Limits
What the report does not hand you
It is prose in a read-only field with a copy button. There is no dashboard, no chart and no export. Rival tools in this category return a tone taxonomy instead: one advertises up to 28 emotional tones, a sentence by sentence breakdown and a confidence percentage with a stated rule that anything over 70 percent is a strong signal.
None of that exists here. No tone list, so no category is assigned. No number of any kind, because nothing on this route computes one. No per-sentence markup, because the route never segments your text. The output is also capped at 1,400 tokens, which is why a very long paste comes back as a report on the passages that matter rather than full coverage.
There is no browser extension, no Gmail or Outlook integration, no API and no batch mode. One draft at a time, in this page.
Cost
What a tone miss costs downstream
A tone problem does not bounce back as an error. It comes back as a slower reply, a question you already answered, or a thread that gets forwarded to someone more senior than it needed to reach.
The hedged recommendation is the expensive one. Nobody can act on it, so it turns into a meeting. The cheerful sign off on bad news is the second, because it makes the sender look like they have not understood the situation. Both pass every correctness check you own.
Straight answer
Does the Claude tone analyzer detect AI writing?
No. It is instructed never to claim certainty about authorship or about which model produced a text, and to describe observable signals and limitations instead. A tone finding is a property of the writing, not evidence about who wrote it.
If you came here from a search for a Claude detector, this is a different tool answering a different question. We make no prediction about what any detector will return and have no affiliation with Anthropic.
In practice
Who puts this in their workflow
Support teams running drafted replies past a check before they go out. Managers writing feedback they cannot take back. Freelancers and agencies sending client updates where sounding uncertain costs work. Anyone whose outgoing writing starts in a model and finishes in their name.
The step that makes it stick is narrow: one run, on the draft that matters, before the send, then edit only what the report names. Rewriting the whole message defeats the point, because the message was already correct.