brand-voice

Gather 5 to 20 representative samples of real writing such as posts, essays, or emails, then have the Skill extract rhythm, compression, claim style, and banned moves into a structured VOICE PROFILE. That profile is reused across all later writing tasks instead of re-deriving style each time.

The strongest sources in priority order are recent original posts and threads, then articles, essays, and newsletters, then real outbound emails that worked, then product docs and site copy. Generic platform exemplars should never be used as source material.

Yes, the profile includes channel notes for X, LinkedIn, and email so the same core voice adapts per platform. Downstream skills like content-engine and crosspost consume the profile directly for consistent output.

Hard bans include fake curiosity hooks, "not X, just Y" constructions, "no fluff", forced lowercase, LinkedIn thought-leader cadence, bait questions, "Excited to share", and generic founder-journey filler. Any of these found in output are deleted and rewritten.

Avoid it when the source set is too thin or unrepresentative, since a weak profile produces misleading mimicry. Also do not persist personal voice fingerprints into repo-tracked files unless the user explicitly requests a durable artifact.