Skip to main content
Yuzu generates buyer-facing material: recap scripts, emails, LinkedIn posts, business-case copy, titles, and summaries. Content quality is the system that keeps that output specific, useful, and less generic.

​
Model routing

Different tasks use different model tiers:
  • Opus for deeper agent work, long-context reasoning, and high-stakes judgment.
  • Sonnet for high-volume generation, summaries, emails, and business-copy drafting.
  • Haiku for fast classification, routing, dedupe, and other lightweight decisions.
The model registry gives Yuzu one place to route work and audit which model generated a result.

​
Judge passes

Some generated content can run through an LLM judge panel. Judges score the draft against a rubric and return concrete feedback. If the first draft is weak, Yuzu can revise once using the judge feedback, then compare the revision against the original. The system keeps the better version and avoids exposing judge language to buyers.

​
What judges look for

Rubrics vary by surface, but generally check:
  • Specificity to the buyer and deal
  • Clear next step
  • Factual grounding in available evidence
  • Avoidance of generic AI phrasing
  • Appropriate tone for a seller-authored artifact
  • No invented claims beyond the source context

​
What still needs human review

Judges are a quality gate, not legal or commercial approval. Review anything that mentions pricing, commitments, timelines, discounts, security, or customer claims before sending.