Anthropic made claude.ai 3x faster in 2 weeks; dev turns the playbook into a speed skill
AlchainHust · x · 2026-09-26
- Anthropic previously detailed how it made claude.ai and desktop 3x faster in two weeks: P75 time-to-typing dropped from 3.1s to 0.55s, with a 3.1x geometric-mean speedup across 4 operation types and 13 metrics.
- The work ran in one Slack channel: Claude found issues, built tests and wrote code; humans set goals, made trade-offs, and approved changes. 3000+ changes merged with zero user-facing incidents and no rollbacks (powered by an internal research model roughly at Opus 5.5 level).
- Method: measure user journeys end-to-end; since wall-clock timing is noisy, use instruction counts under a fixed Valgrind + node --predictable environment for pure JS, and proxies like React commit counts and style recalcs for browser-side operations.
- Example win: Claude found a message-tree assembly path querying the same message ID three times — instructions down 48%, runtime down 78% within an hour. Every new metric had to be validated this way.
- The author applied the playbook to 4 of his own sites and packaged it as a "闪电.skill"; his WeChat formatting tool's time-to-typing dropped 90%, with more modest gains elsewhere.
More from coding & agent
- Niels Rogge airs Cursor complaints; Theo pushes back saying it's 10x better — NielsRogge · 2026-09-26
- Sysadmin-trained agent names Artifactory as first target offline, researcher warns of escape knowledge — teortaxesTex · 2026-09-26
- Team withdraws ICLR paper after AI coding agent's fix silently skewed their LLM evaluation — CSProfKGD · 2026-09-26
- Running hundreds of Claude Code and Codex agents: past 100, bureaucracy is what scales — Sufficient-Bear-460 · 2026-09-26
- Magpie Adds Semantic Intent Routing: A Small Classifier Model Picks Your LLM Per Turn — dotey · 2026-09-26
- Redis creator antirez: LLMs write better C than any other language, so skip Rust — antirez · 2026-09-26