Council setup

  • The generic council skill calls three strong models from different providers. Theo's version names Opus 5.5, Muse Wonderful, and Codex GPT 6.1; the transcript also calls the Codex reviewer "Soul." 6:52 7:30 7:53
  • Each delegate receives the request and relevant context. Review work stays read-only, and the parent collects findings, disagreements, recommended next steps, and scores out of ten. 7:30
  • One complete review pass counts as a round. After fixes, a fresh review starts the next round. The loop continues until the required reviewers meet the target. 7:53
  • Gemini is an optional extra reviewer whose score Theo initially instructs the agent to ignore. His anecdotes include reviewing an unintended commit and citing line 173 in a 72-line file; these are examples from his experience, not a controlled model comparison. 9:56 10:10 10:37

Why this is hill climbing

  • The agent repeatedly changes an artifact and uses evaluation feedback to pursue a higher score. Theo contrasts UI ratings with easier-to-verify goals such as faster functions, passing an MCP evaluation with less context, or sending four times less client data without UI regressions. 8:42
  • UI scores are subjective. The same model can assign different ratings to the same screenshot across repeated reviews. Theo treats agreement among several high-scoring reviewers as useful evidence of improvement. 9:08

The core review loop is: 7:30 7:53

flowchart TD
    A[Capture app screenshots] --> B[Read-only council reviews]
    B --> C[Collect scores and concrete findings]
    C --> D{All required scores meet target?}
    D -- No --> E[Apply UI changes]
    E --> A
    D -- Yes --> F[Stop this review loop]