What's this about?
Claude profiled paths, proposed changes, built benchmarks and watched deployments while humans chose goals and approved every merge.
What is a performance ratchet?
After an optimisation, a benchmark threshold is tightened so later code cannot silently give the gain back. It turns a one-off speedup into a permanent CI guardrail.
Why does this matter?
launch post 5,900-plus X likes.
- Fresh-page typeability fell from 3.1 seconds to 0.55 at the 75th percentile.
- The team rejected metrics that were flaky or unrelated to user latency.
- Long conversations and 95th-percentile performance still need work.