better_bytes.md

better_bytes.md

claude by The numbers:

• 55.8% of your signals are taste — you giving research direction
• 20.8% interrupts — Claude going wrong way, you cutting it off
• 17.4% approvals — Claude running autonomously and you saying "keep going"
• 6.0% explicit redirects — "no, try this instead"
• 87.1% self-investigation ratio — when Claude faces a choice, it decides rather than asking (only 9 unnecessary asks)

  Opus 4.5 → Opus 4.6:

  ┌──────────────────┬──────────┬──────────┬────────────────────────────────────┐
  │      Metric      │ Opus 4.5 │ Opus 4.6 │             Direction              │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Interrupt rate   │ 46.1%    │ 54.8%    │ worse — you're cutting it off more │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Redirect rate    │ 4.9%     │ 9.5%     │ worse — you're correcting it more  │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Approval rate    │ 31.4%    │ 22.2%    │ worse — fewer "keep going"s        │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Taste dependency │ 17.6%    │ 13.5%    │ better — needs less direction      │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Self-correction  │ 0.73%    │ 0.54%    │ worse — catches fewer own mistakes │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Unnecessary asks │ 0        │ 4        │ slightly worse                     │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Max run length   │ 6        │ 9        │ better — longer autonomous streaks │
  ├──────────────────┼──────────┼──────────┼────────────────────────────────────┤
  │ Long runs (5+)   │ 26       │ 38       │ better — more long autonomous runs │
  └──────────────────┴──────────┴──────────┴────────────────────────────────────┘

  But caveat: Opus 4.5 is 33 conversations vs 603 for Opus 4.6. The Haiku numbers are meaningless (497 automated batch conversations with no human interaction — that's your autointerp
  pipeline, not real dialogue).

  Sonnet 4 only has 5 conversations so too few to read much into, but 2.31% self-correction rate is notably higher.

  The headline: Opus 4.6 does longer autonomous runs but you interrupt and redirect it more. It's more aggressive but less aligned with your intent. The taste dependency dropping (17.6% →
  13.5%) suggests it's more opinionated — just not always in the right direction.