The job
What to use the Claude detector for
One use of this works and one does not. Checking your own draft works: you get the tells named, you look at the text, you edit. Judging someone else's work does not work, because nothing this tool produces supports a decision about another person, and no detector's output does either.
Mechanically it is simple. Paste a passage, press Analyze, and a written reading comes back. It runs on our server, 5 runs per rolling 24 hours, up to 12,000 characters each, with an optional instruction of up to 300 characters. Nothing is stored, so the Copy button is how the findings leave the page.
The loop
How to check a draft in three steps
This is an editing loop, so run it at the point where you can still change things.
- Step 1
- Paste one section rather than a whole document. Findings about four paragraphs are actionable; findings about four thousand words are a summary.
- Step 2
- Name your suspicion in the instruction box. "Is this hedged?" or "Do all these paragraphs open the same way?" gets a sharper answer than nothing.
- Step 3
- Test each finding against the text and edit the ones that hold. Then run the next section, not the same one again.
The edit list
Four things to strip out of a draft
These come from the signal lists that two detector pages publish about Claude text. As editing instructions they are useful regardless of how the draft was written.
- Hedges in front of claims
- "It is worth noting", "generally", "in many cases", "it seems", "tends to". Delete the qualifier and see whether you still believe the sentence.
- Openers that warm up
- A qualifying clause before the point. Move the point to the front of the sentence and cut what is left.
- Structure you did not need
- Numbered steps, headers and tidy summaries applied to material that had no steps in it. Keep the ones doing work.
- Perfect evenness
- One page describes the tell as consistently high readability throughout. Let one sentence be short and blunt and the paragraph stops humming.
Set expectations
Why the Claude detector gives no verdict
You will not get a percentage, a confidence score or a set of highlighted sentences here, which means there is nothing to paste into an accusation. That is the design.
The detectors that do produce those things say the awkward part in their own copy. Plagly.ai, after listing five Claude tells, writes: "None of these is proof on its own, and careful human writers legitimately share several of them." AIDetego, on a page selling a Claude detector, states that "Exact attribution to a specific model (Claude vs. ChatGPT) is currently not possible" and that "no tool can reliably do this."
A separate question
Watermarks are not what this checks
A page at overchat.ai claims that Claude models add an invisible statistical pattern to their output by influencing token selection during generation. We cannot verify a claim about another company's models, so we neither repeat it as fact nor deny it, and this route detects nothing of the kind.
The same page is useful for a different reason. Its own scanner looks for zero-width characters, hidden Unicode, punctuation style and formatting residue, and it says: "Finding one of these artifacts therefore does not prove that text was written by Claude, and removing them does not necessarily remove Claude's watermark." Two separate things, two separate tools, and neither of them is this one.
Spending the allowance
Five runs, and how to use them on one piece
The window is rolling rather than daily, so runs spent this morning return tomorrow morning. On a long piece that is enough for a real pass if you aim them at different sections rather than re-running the same paragraph hoping for a different reading.
A workable split: the opening, the section you wrote last and fastest, and the section you are quietly proud of, which is often the most over-written. That leaves two in reserve for after you have edited, which is the only time a second look tells you anything new.
Where it stops
What this Claude detector does not do
No verdict, no score, no bands, no highlighted sentences, no model attribution and no accuracy figure. No watermark detection and no hidden character scan on this route.
No file upload, no unlimited scans, no API, no account and no stored history. And no claim about what any other detector will report about the same text, because we cannot see their output and would not predict it if we could.
One competing page advertises unlimited free scans, another publishes a 99% accuracy figure, and a third names five languages it supports. We match none of those and list our own numbers instead, because a figure nobody here has measured is worth less than an honest blank.