@di-atomic/voice-builder · v0.3.2 · beta

Your agent doesn't need a voice. It needs the numbers behind one.

voice-builder turns your brand voice into a specification a machine can be held to: reading ease, sentence rhythm, hype density, direct-address density. Then every downstream skill writes to those numbers instead of guessing.

For teams whose AI writes customer-facing copy. Replaces the "professional yet approachable" paragraph that every brand gets.

15starter voices you install instead of inventing
7languages carried by transcreation, not translation
4scripts that gate the output with numbers
100/100marketplace security score · tier beta

Install

Guidance-only: there is no backend and no API key. Your agent reads the skill and does the work.

# OPVS dashboard → Marketplace → @di-atomic/voice-builder → Install # then, in any agent session, just say: "build my brand voice"

What you actually get

📚

A starting voice in one move

Install the closest of 15 starter voices (4 traits × 4 ICPs × 7 languages) and tune it, instead of re-deriving a generic voice on every single call.

📐

Targets your copy can fail

You get measured numbers, not adjectives. Downstream drafts are checked against them, so "off-brand" becomes a measurement rather than an opinion.

⚖️

A gate that holds for regulated brands

Selling biocides or CLP-classified products? A playful voice is refused outright and every claim is linted against BPR Article 72.

The difference, on my own brand

This is the real Di-Atomic voice profile, produced by this skill. Left: what a describe-the-tone prompt returns. Right: what voice-builder writes.

Description — nothing can be held to it

"Professional yet approachable. Clear, confident and customer-focused. Speaks with authority while remaining friendly and accessible."

Specification — every line is checkable
  • Registry match icp-agency 0.93
  • Flesch reading ease 59.9
  • Grade level 7.8
  • Second-person / 1k words 53.9
  • Hype density / 1k 0.0
  • Off-brand floor (SaaS jargon) 14.5

How it fits the rest of your stack

voice-builder is a foundation skill. It writes one file, and every content skill downstream reads it.

A gold voice.md document on the left feeding four arrows out to copy-engine, post-writer, outbound-engine and section-designer
One source file, four consumers. Get it wrong and all four are confidently off-brand.

Watch it catch its own mistake

A real trace on a real client, ONYX Radiance. The classifier was confident and wrong, the test caught it, and the fix went into the skill.

$ analyze-voice.py onyx-samples/ median sentence 8.1 words # short sentences. B2B SaaS also writes short sentences. $ match-archetype.py --profile onyx.json trait-terse x icp-b2b-saas confidence 0.93 # confident. also completely wrong — ONYX sells face serum to consumers. # the fix: measure WHO the sentence points at, not just how long it is $ analyze-voice.py --direct-address onyx-samples/ second-person / 1k words 94 $ match-archetype.py --profile onyx.json trait-warm x icp-ecommerce confidence 1.00

Works with

✍️

copy-engine

Writes landing, pricing and email copy against the numbers voice-builder set.

📣

post-writer

Adapts the same voice to each platform without drifting into generic social tone.

📧

outbound-engine

Keeps cold sequences sounding like you, not like a template.

🎨

section-designer

Takes the voice into page sections so the design and the words agree.

Stop letting your AI guess how you sound.

Install voice-builder, spend five minutes tuning a starter voice, and give every downstream skill a target it can actually hit. Free, tier beta, on the OPVS marketplace.