Writing that preserves the person and the facts.

The problem

The common failure is not that AI writing is grammatically wrong — it rarely is. It is that it drifts toward the statistical average: the same handful of constructions (the em dash, "it's not X, it's Y," the rule of three), the same vocabulary (leverage, unlock, seamless), the same shape, regardless of who supposedly wrote it or what they were actually trying to say. Read enough of it and every piece starts to sound like the same anonymous author.

Why not just ask for it in the prompt

A blunt instruction like "make this sound more human" tends to produce a second kind of failure: the words change but the substance does not survive — a caveat gets smoothed away, a specific number gets rounded into a vaguer claim, or the writing becomes performatively casual without becoming any more true. Humaniser exists because "sound human" and "stay accurate" are two different jobs, and doing them in the wrong order breaks the second one.

How it works

• —Applies seven ordered priorities, and a higher one always wins a conflict: preserve truth and supplied facts; preserve the user’s purpose and intended meaning; preserve material context and detail; fit the actual audience, platform, genre and length; match the user’s voice where a genuine sample exists; remove generic AI-writing patterns; then improve clarity and rhythm without inventing evidence.
• —Keeps a silent context-retention ledger of names, numbers, dates, qualifications, caveats and the intended conclusion, and will not trim a ledger item just to make the writing punchier.
• —Runs two internal passes (purpose and retention, then AI-pattern removal) and stops there unless a forensic audit is explicitly requested.

Evidence and provenance

• —Technical, legal, regulatory and scientific writing keep accuracy and genre convention ahead of anti-slop aesthetics.
• —A roughly forty-item vocabulary watch-list (delve, leverage, unlock, seamless, game-changer and similar) is treated as guidance, not a blind blacklist.

What it produces

• —One finished, purpose-fit piece of writing — no explanation of the rules applied unless a diagnostic is requested.
• —On request, a 0–10 "AI smell" score with the specific trigger phrases and the rule each one broke.

What it must not do

• —Never invents a number, quote, source, personal experience or result to make writing sound more specific.
• —Fails its own brief if the output sounds clean but loses an important fact, changes the intended meaning, or turns mixed evidence into false certainty.
• —Never copies distinctive sentences from a voice sample — it learns the pattern, not the wording.

In practice

Take a typical AI-drafted line: "Our platform leverages a robust, seamless architecture to unlock transformative growth for your business." It is grammatical and says almost nothing. A surface "humanising" pass might only swap the vocabulary, producing "Our platform uses a strong, smooth setup to create real growth," which still says almost nothing, just in plainer words. Humaniser's first move is different: it asks what the sentence was actually trying to claim, finds there is no real content behind it, and either replaces it with the specific claim the writer meant ("the system reduced average handling time by keeping context in one place") or flags that the sentence has no real content to preserve.

Why a reusable skill, not a one-off prompt

Anti-slop rules re-typed into every prompt drift over time: a rule remembered today is a rule forgotten under deadline pressure next week. Packaging the seven-priority order and the retention ledger as a skill means the same discipline (truth first, style last) applies consistently, across every piece of writing, not just the ones where someone remembered to ask for it.

Mohammed's contribution

Defined the seven-level priority order and, specifically, its governing rule that a higher priority always overrides a lower one in a conflict — the mechanism that keeps a "more human-sounding" rewrite from being allowed to quietly change what was actually said.