The pack as data a script can hold a draft against
Your team already has most of this pack — scattered across a tone doc from onboarding, a "please stop promising refunds" thread in the team channel, and whatever the KB currently says. The pack's job is to gather it into one structure with three properties: it's testable, it's versioned, and a script can read it. Run the code above — that's a real pack, minus the parts specific to your product.
Walk the fields:
tone— each rule is checkable against a draft, not a mood. "Apologize once, then act" is testable: count the apologies, find the action. "Be warm" is not testable, so it's not a rule; it's a wish. If a new hire couldn't look at a draft and say pass or fail, rewrite the rule until they could.banned_promises— the load-bearing list. Every phrase on it is something a drafter has no authority to say: guarantees you don't control ("we guarantee", "arrive by" — the carrier decides that, not you), money you can't move ("full refund" — refund authority lives with a human, always), exceptions you can't grant, and futures you can't see ("won't happen again"). These are phrases, not concepts, on purpose: a script can substring-match a phrase. It cannot substring-match a concept.kb_index— every article your drafts may cite, each with alast_verifieddate. Notice what the run just told you: KB-204 is STALE. Nobody has confirmed the shipping-delay article since November, and the carrier changed its status names in spring. Any draft citing it should stop at the gate until a human re-verifies the article and bumps the date. An unverified article is not a citation target — it's failure mode #3 from the last step, pre-loaded.version,owner,reviewed— the pack is maintained like code because it is code, in the sense that matters: things downstream break silently when it rots. When the return window changes from 30 days to 21, one person updates one file and every future draft gets checked against the new truth. When it lives in six heads, the six heads disagree by August.
Why the QJE study is really about the pack
The 5,200-agent study you'll meet properly in a few steps found something subtle: part of the AI assistant's productivity gain came from disseminating the practices of the strongest agents to everyone else. The assistant had, in effect, learned the best agents' moves and was suggesting them queue-wide.
A policy pack is that same dissemination, done on purpose and under version control. Your best agent's instincts — what never to promise, which article is actually current, when to stop typing and escalate — written down where the drafter, the script, and the newest hire on the team all read the same copy. That's the artifact. The next six steps teach the checking.