01
Hypothesis
Voice consistency is the hardest problem in AI copy production. Every agent has its own style. Every session resets. Every page has a different format. The question: can a single reference file anchor the voice of 10 different agents, producing copy that sounds like the same person across 9 pages?
The risk was real: radically different styles, from Marshall's logical structure to Urban's narrative recursion and Halbert's direct response. Forcing them to converge without losing what each does uniquely.
02
Method
A 145-line voice file with two layers: the tone of direct conversations and the brand tone of public content. Power words, approved metaphors and explicit anti-patterns (Instagram coach, corporate, hype). The C1 (voice) criterion works as an automatic fail: below 0.60, the page is rejected regardless of everything else.
Every page went through COPY FORGE with this file as required context. Agents had stylistic freedom within the file's limits.
03
Result
C1 = 5/5 on 83% of pages (5 pages, 8 different agents). C1 ≥ 4/5 on 100%. Zero serious drift. The agents orbit different styles, but the voice converges on the same center.
The key point: the file does not prescribe sentences. It prescribes limits. The 7 "never" lines were more effective than the 20+ "always" lines.
04
What changed
The voice file became infrastructure, not a document. It is loaded into every content and copy pipeline. The "limits over prescriptions" pattern was validated: defining what the voice is not produces more robust consistency than defining what it is.