A pack that can defend itself
Run the editor. That's the finished artifact: measurements turned into rules, plus three fields that most brand books don't have and every voice pack needs.
derived_from. The pack names its evidence. When the client's new marketing lead — there's always a new marketing lead — says "this doesn't feel like us," you don't argue taste against taste. You point at three pieces of copy their own company published and approved and the measurements that came out of them. "Cap at 15" stops being your opinion and becomes their about page's opinion. That argument ends differently than the vibes argument does.
version and owner. Voice packs change — a rebrand, a new product line, a legal scrub of the superlatives. When that happens, the version bumps and every check downstream picks up the new rules on the next run. What you never do is edit silently: a batch checked against v1 and a batch checked against v2 are different claims about what "on-voice" meant that day, and three weeks from now you may need to know which was which. If versioning-like-code sounds heavy for a writing job, wait until lesson 3, where the same habit is the difference between a correction and an incident.
banned. The one list that isn't derived from the samples — it's derived from the model. "Game-changing", "unleash", "elevate": you already know the sludge that generative drafts reach for, because you've deleted it a hundred times. The banned list is that deletion, written down once, enforced forever. It'll grow. Every time a draft makes you wince, the wince goes in the list, and the list outlives the draft.
And reader — one line on who's actually reading. "Packs light, reads specs, allergic to hype" does more work in a drafting prompt than three paragraphs of brand poetry, because it's falsifiable: a line either respects a spec-reader or it doesn't.
Two jobs, one artifact
The pack works both directions. Going in, it rides along in the drafting prompt so the model starts near the voice. Coming out, the same dict drives the mechanical check on every variant — because a prompt is a polite request, and across a twelve-variant batch, polite requests leak. You'll build that outbound check in the next lesson. First: prove you can fingerprint samples yourself.