Reference: Brand-Voice Content Generation

Architectural notes

Reference architecture for a content team's brand-voice agent: drafts
blog posts, social copy, and emails in the company's voice with
source-grounded claims.

Why these picks:
- Brand-voice samples stored in turbopuffer as a per-team namespace.
- Zep tracks rejected drafts as negative training signal.
- Latitude for the brand-voice prompt registry — versioned, A/B tested.
- Notion MCP because most content briefs live there.
- Eval against editor-approved golden drafts catches style regressions.

Common gotchas:
- Generic LLM output is the failure mode; invest hard in voice samples + style guides.
- Always require a citation pass before fact claims ship.