Fingerprint-calibrated voice
I calibrated each voice fingerprint to a product’s measured baseline, so voice became a target instead of a feeling.
AI content governance · Meta · 2025–2026
Verso
Good guidance is written for people. Agents need a checkable target.
Meta’s content design org was moving from writing every string to governing what agents wrote.
- Every week another content designer had built their own agent: a prompt file shaped like software, and a prompt file shaped like software fails like software. Quietly, with nothing built in to say so.
- Review boards were the industry’s answer to agent output, at Meta as everywhere: people reading drafts and deciding whether they met the bar. It was the responsible move with the tools that existed. But a review board is a linter made of humans: slow, and no more consistent than the thing it’s checking.
I built Verso to give agents the target while keeping the kinds of authority separate.
- Role
- Creator and builder of the entire system
- Team
- Feedback from content designers · partnered with central content design to adopt it
- Timeline
- 2025–2026
- Surface
- Meta’s internal content tooling, company-wide
What made Verso different
A content design agent that thinks like a coworker, not a linter.
Most AI writing helpers grade grammar. I built Verso to do content design: reason about user state, surface, product pattern, and evidence, then say what it would ship and why. Words are UI. Bad content is a bug.
Design insight
Voice was written in adjectives. I turned it into a target an agent could be held to. Guidance such as “concise” and “genuine” described what a reader feels, not what a writer or agent does.
| Dimension | Typical writing AI | Verso |
|---|---|---|
| Unit of work | Sentences & grammar | Treats product strings as design decisions. |
| Authority | Generic style rules | Grounds recommendations in governed product sources and shipped patterns. |
| Output | A corrected paragraph | Returns a recommendation, evidence, and tradeoff. |
| Consistency | Per-document | Reuses patterns across surfaces. |
| Voice model | One tone slider | Calibrates to a ten-dimension fingerprint built from product baselines. |
| Personality | Neutral assistant | Has enough personality to say, “that’s a paragraph trying to escape.” |
How it worked
I built Verso to bring judgment to the work, not another score to decode.
-
It’s a teammate, not a tool.
I designed Verso to answer like a senior content designer pulled into a messy thread. It gave an opinion, the reason behind it, and a next move, not a score to decode.
-
Opinions backed by evidence, not vibes.
I grounded every recommendation in governed guidance, product patterns, or shipped strings. When the evidence ran out, Verso said so instead of bluffing.
-
It runs real checks before it opens its mouth.
I built the ten-dimension engine to test voice, directness, rhythm, density, and other language signals. Designers got the human translation, not raw JSON.
-
It knows the difference between jamming and reviewing.
I made its posture match the task. For brainstorming, it explored. For an audit, it structured the work, cited its evidence, and showed its reasoning.
-
It gates the risky move, never the useful work.
I let Verso draft, research, and analyze freely. I gated only consequential actions such as sending, publishing, or shipping, with the artifact already prepared for approval.
What was genuinely new
I made product judgment reusable without pretending it was a rule.
Pattern-first instinct
I made it look for a proven shipped pattern before inventing a new one. Consistency beat cleverness.
Source-privacy awareness
I designed it to distinguish private notes from team guidance and not quietly carry private content into a wider audience.
Right-sized effort
I sized the work to the request: a bounded review for one string, the full machinery for an audit, and no research project for a gut check.
The bottom line
I treated content as a design surface and built Verso to run the checks, remember patterns, cite sources, and still exercise the judgment to say what it would ship.
