Voice Builder

I shipped the first version of this skill and rated it, in my own notes, "not useful at all." It read three samples, described the tone back to me, and produced a voice.md that every brand could have shared. Describing a voice is not the same as specifying one. voice-builder now ships 15 starter voices: 4 trait variants (terse, warm, authoritative, playful, modeled on Stripe, Mailchimp, IBM and Duolingo), 4 ICP archetypes, and 7 language profiles. A starting voice is trait x ICP x language, so the agent installs the closest one and tunes it instead of guessing fresh on every call. Four scripts do the measuring: readability, archetype-match confidence, multilingual fidelity, and an EU compliance gate. It caught a real bug on a real client. The classifier scored a skincare brand as B2B SaaS at 0.93 confidence on sentence length alone. Adding direct-address density (94 second-person words per 1,000) re-landed it correctly at 1.0. Everything downstream reads what it writes: copy-engine, post-writer, outbound-engine, section-designer.