Write in Your Voice

Version 1.1.0. Free. Works with major AI agents. Source: github.com/diaryofamademan/skills

What it does

A voice cannot be learned from text the person never wrote or said. Most voice profiles are built exactly that way, from the person's published posts and a paragraph describing their style, and they fail in a specific way: if any of those posts were ghostwritten or AI-generated, the model learns its own voice back and returns it as theirs, and the failure is invisible because the output reads fine. This skill builds the evidence base first. It assembles a source bank from how the person actually speaks, their call and video transcripts, voice notes, and dictations, layered in a fixed order with published posts quarantined at the bottom as contaminated. Drafts are composed from the person's own sentences, never generated first and voice-checked after. Where no speech covers the move, an approved text may source the draft, and the run has to declare that it did; where nothing in the bank covers it, the skill returns a dictation prompt instead of a draft. Two gates check the result: a word gate on vocabulary, and a shape gate on the draft's form, because a draft can pass every word check and still not sound like the person who has to sign it.

The skill ships with its proof: a production source bank of 690 passages cut from 1,149 transcripts, about 224,000 words of speech, with the published posts quarantined as contaminated, and one launch written two ways, where the vocabulary checker scored the rejected draft cleaner than the approved one, 2 flags against 3. Read the launch-pair case study on GitHub

Use it when:

  • You are setting up AI drafting for a founder, creator, or executive, and it has to sound like the person from the first draft, not like a model describing them.
  • Drafts keep coming back with "not my voice," even after they pass every vocabulary and style check you run.
  • A voice profile built from published posts produces plausible text the person rejects on sight, and you cannot see why.

What is inside:

  • The two procedures: build the source bank (gather speech, purge other voices, cut and score passages, layer the four tiers under a header that records the person's channel conventions, wire the bank into drafting, maintain it against a kill-rate ledger), and draft from the bank (source first, compose from their sentences, approved texts may source where speech has no coverage, declared in the run, and no coverage anywhere means a dictation prompt).
  • The gates in detail: the word gate, built from the person's spoken vocabulary plus a never-use list, with every flag rewritten away or justified in the run's return, and the shape gate, with its parallelism scan, press-release test, and read-aloud check against anchors chosen for the piece type, plus the rule that the voice outranks any house register.
  • The reference files: the source bank process step by step, the drafting procedure and the gates in detail, and the worked example of the production build the skill was learned from, with its failure left in.

Get the skill

Put your email in. You get the install link and a one page quick start right away, and you join the free newsletter for updates.

Your email is used to send the skill you asked for and the free newsletter. The list runs on Beehiiv. You can unsubscribe in one click in any email. Your address is not sold. If you apply for the audit, that information is used only to review your application, unless you also tick the newsletter box on that form.