Machine copy is everywhere. Sounding like a person is the scarce thing.

Every channel is filling with machine writing, and readers have learnt the tells. The stock phrases, the tidy rule-of-three paragraphs, the summary nobody asked for. The strongest tell is over-connection: every point wired to the next, every clause resolved, nothing left for the reader to close. People do not write like that. And you cannot fix it by asking the model to sound natural, because the model's own idea of natural is exactly what produced the tells. As the copy piles up, language is quietly becoming one of the biggest differentiators there is. We wanted the writing under our name to hold up, so we built the toolkit.

We sent 14 research agents through the published literature and came back with 462 findings, 60 of them adversarially verified against their sources. That became a working corpus: a 116-entry blacklist of machine-tell phrases, 278 structural rules, 335 hard statistics, 85 industry-specific rules. Sources span UK style guides (GOV.UK, the Plain English Campaign, the Economist, the FT, the Guardian, the Times), peer-reviewed linguistics, the AI-detection literature, and UK workplace surveys with real sample sizes.

The load-bearing anchors are named and checkable. Biber's register dimensions, which measure whether prose sits at the involved or the informational pole, so an email is not edited like a solicitor's letter. Burrows's Delta, the authorship statistic showing each writer has a stable function-word fingerprint, which makes the author's own idiolect the target rather than some generic informal voice. The Appraisal framework, for restoring the writer's stance where a draft is a stack of bare assertions. Perplexity and burstiness, the token mechanism behind detectors: models pick predictable words in an even rhythm, and breaking the connections and varying the rhythm dissolves that at the source. And the human-perception studies (one, in PNAS, ran 4,600 readers) showing most folk cues for spotting AI are non-diagnostic or backwards, which is why the toolkit removes tells rather than injecting fake quirks. Real samples of the owner's own typing anchor the register, so edits move toward a person, not a persona. None of this ships inside the models themselves.

The machine writes the first draft. The draft goes through the pass: blacklist hits out, structure given the variation a human editor would give it, register tuned against the corpus and held to the target, plain British English or native Polish, in both directions across that language line. Then a person reads it and signs it off. Nothing ships on a score. The same discipline runs on video: the clip engine carries roughly 21 quality gates, voice register included, so the delivery stays flat and unhyped before anything renders.

The toolkit's first client was this website. Once the rules existed, we turned them on the pages that sell them: a full pass over the finished site caught the machine tics the drafting had left behind, the same polished phrase recycled across pages, paragraphs wired too neatly end to end, and cut them, this page included. The blacklist was calibrated the same way, against a real inbox rather than theory: 140 machine-labelled spam emails set against 66 genuine human ones taught the scorer what the human pile actually reads like. The clip gates came out of the same habit. Build the rule, run it on your own work first, keep what survives.

You cannot buy the toolkit on its own, because everything cooltake ships already goes through it: the naturalness pass comes with every tier, pages, emails, outreach, scripts and clips alike. If your copy has started reading like everyone else's, that page is the place to start.