1. Write limits, not sentences
The voice file in the experiment had 145 lines. The 7 "never" lines held the voice better than the 20+ "always" lines. Banning the coach tone, corporate jargon and hype leaves room for each agent to write its own way without leaving the brand.
2. Split tone by context
The same file had two layers: the tone of direct conversations, more colloquial, and the brand tone of public content. Without that split, a support agent sounds like a manifesto, or a landing page sounds like a text message.
3. Load it into every agent
The file goes in as required context in every pipeline that writes. A new agent gets the file before writing its first line. A voice that depends on remembering does not last.
4. Give voice a score that fails on its own
In the experiment, voice had its own criterion (C1) inside the quality score. Below 0.60, the page was rejected regardless of the other criteria. That stops a well-structured text from passing with the wrong voice.
5. Share a narrative DNA
Voice makes each page sound the same. For pages to connect with each other, experiment #004 used a second 139-line document with the core concepts, 7 approved analogies and one rule: each page uses one line from the system, never all of them.
The result
| Measure | Result |
|---|---|
| Agents writing | 10 |
| Pages | 9 |
| Voice 5/5 | 83% of pages |
| Voice 4/5 or higher | 100% of pages |
Frequently asked questions
Do I need text samples in the voice file?
They help as reference, but limits are what hold the voice. Too many samples make agents copy sentences.
Who scores the voice?
An evaluator agent with written criteria, separate from the writer. It is an isolated model judge, the third level of verification.
Want to apply this to your operation?
The first call is a free one-hour diagnosis.