Resetting the reference point

  • Theo starts a fresh thread with the existing PR, asks reviewers to generate screenshots, and requests another PR stacked on top. He then adds Claude Code and Codex screenshots, asking for T3 Code to score consistently higher than both. 13:45 14:08 14:28
  • With comparative context, Soul rates main at 7/10 and the previously approved PR at 7.4. The earlier 9/10 does not survive the new framing. 15:05
  • Muse rates main at 6.2 and the updated UI at 6.9, versus 8.1 for Claude Code and 7.6 for Codex. Reviewers disagree about both absolute scores and product rankings. 16:35

Controlling usage and improving consistency

  • To conserve usage, Theo gates reviewers: iterate with Soul first, bring in Opus after Soul scores high enough, then bring in Muse after Opus improves. Soul progresses through 7.4, 8.1, 8.9, and 9. 17:01 17:23
  • Opus progresses from 7.1 to 8.7 and eventually 9, matching Soul. Muse still gives only 7.9 before a usage error stops this stage. This stage therefore does not end with unanimous 9/10 approval. 17:23
  • Concrete gains include consistent labels, a less visually dominant sidebar, more consistent fonts, radii, and indicators, and better body-text contrast, especially in light mode. 17:23 17:52
  • Scores for unchanged competitor UIs rise alongside T3 Code's score. Comparative screenshots help expose weaknesses, but they do not make the reviewers' numerical scale stable. 15:31 15:43