Voice Trainer

The single biggest tell that content is AI-generated isn't a specific word — it's the absence of a human voice. Real writers have patterns: sentence rhythms they return to, types of humour they lean on, phrases they overuse in a charming way. AI output is smooth, even, and forgettable. A voice profile fixes this — once captured, it becomes a reference file that transforms generic AI output into something that sounds like a person wrote it.

The signal check comes first because quantity isn't signal. Three samples of corporate boilerplate ("We are pleased to announce…") contain almost no voice to extract — those are genre conventions, not a person. If the samples are near-voiceless or all in one flat register, the skill says so and stops rather than manufacturing a personality the evidence doesn't support. A falsely flattering profile built from press-release filler ("enthusiastic, collaborative, customer-focused") is worse than no profile because the user will trust it. Instead, the skill names the personality-forward sources it needs: casual emails, texts, Slack messages, unedited drafts, transcribed voice memos.

Extraction runs three passes — mechanics, personality, structure — and the structural pass is where the value sits. The "transferable move" is the structural thing the writer does that's independent of subject: "opens on a concrete sensory detail then pivots to a self-aware insight" or "states a blunt claim then immediately undercuts it with a parenthetical". A recurring topic (coffee, running) is not a voice pattern; the skill looks past topic to the structural move underneath and encodes that. The output includes three example transformations, and at least one must be on a topic the samples never touched — proof the profile captured transferable style rather than the writer's recurring subjects.

Two rules guard the output. Capture, don't correct — even when the user explicitly asks to "make it more confident while you're at it". A cleaned-up profile describes a person who doesn't exist; future content from it sounds less like the user, which is the opposite of the point. And every quoted example is a verbatim substring of a supplied sample — no paraphrasing inside quotation marks, no stitched fragments, no tidied casing. A near-miss quote silently teaches the wrong voice. If no verbatim quote evidences a pattern, the pattern gets dropped or marked tentative — never asserted.