kReative Labs
Back to kReative Labs

Live 2026

unslop-text

Paste a draft in and find out what gives it away as machine-written. Two passes, because the two kinds of tell need completely different machinery to catch.

Try the scanner

The idea

There are two kinds of tell and only one is mechanical

Some giveaways are fixed strings. Particular punctuation, a small set of overused words, a handful of stock constructions. A regular expression catches those perfectly, instantly, and without sending your draft anywhere.

The rest are structural. Rhythm that never varies. Sycophancy. Saying very little at considerable length. Those are cited by readers just as often, sometimes more, and no pattern can detect them because the problem is not which words appear but how the whole thing moves.

So the tool is honest about the split. A clean mechanical scan means the lexical layer is clean. It does not mean the writing reads as human, and the interface says so rather than letting a green score imply otherwise.

Architecture

Local first, model second

unslop-text two-pass architecture Pasted text stays in the browser for pass A, a regex scanner that produces ranked findings and a slop score with no network request. Pass B is optional and sends the draft through the application's own server, which holds the API key, to a model for a structural read. Results from both passes combine into one report. BROWSER · NOTHING LEAVES THIS BOX Your draft pasted in Pass A: regex scan fixed-string tells, ranked by how often readers cite them Findings + slop score severity filters, per-item fix App server holds the API key never the browser Model structural read Pass B: structural audit rhythm, sycophancy, length without content optional
Pass A never touches the network, which matters when the thing being audited is an unpublished draft. Pass B is opt-in, and the key stays server-side so the browser never holds a credential.

Detail

At a glance

Pass ARegex scanner, runs entirely in the tab. No network request, nothing stored.
Pass BOptional model-based structural read, proxied through the app's own server.
Key handlingAPI key server-side only. Never exposed to the browser.
OutputRanked findings with severity filters, a slop score, and a suggested fix per item.
CompanionShares its audit with a Claude skill of the same name.
StageLive.