How we made claude.ai 3x faster in two weeks
3 hours ago
- Claude.ai and desktop app improved about 3x in a two-week sprint by measuring and optimizing user journeys.
- Four key user journeys (95% of activity) were targeted: launching app, starting conversation, loading existing conversation, sending message.
- 75th percentile load times dropped from 3.1s to 0.55s (fresh web load), 0.8s to 0.3s (new Claude Code session), and 2.6s to 0.73s (cloud session).
- Claude used an internal research model (comparable to Opus 5.5) to find bottlenecks, build benchmarks, and ship improvements.
- Over 3,000 changes were merged without a single customer-facing incident or rollback.
- A Slack channel with Claude as facilitator monitored performance, analyzed Datadog data, and proposed optimizations.
- Initial 20 projects hit 12 of 13 targets within three days; Claude then identified additional opportunities exceeding original goals.
- Measurements like instruction counts, V8 call counts, and layout shifts were used as deterministic benchmarks to drive optimization.
- Claude cut instruction counts by 48% and 31% on hot paths, reducing wall-clock time by 78% and 44%.
- Layout instability was addressed by mapping shifts to named regions and adding integration tests, fixing 31% of loads with unwanted movement.
- Claude found and fixed issues: unnecessary React hooks, slow CSS selector, hidden location.reload() causing 500k daily reloads, and UTF-16 slowdown from em dashes.
- Safety mechanisms included automated reviews, human approvals, unit tests, feature flags, and incremental rollouts.
- Over 200 feature flags were introduced and cleaned up; dozens of guardrails protected performance wins.
- A sidequest improved streaming reply performance: reduced main thread blocking from 750ms to 200ms, used 1/3 CPU, maintained 120fps.
- The sprint loop: find measurable metric, Claude climbs it, guardrails prevent regression, all in open Slack threads.
- Human team provided ambition, taste, and direction; Claude autonomously opened new threads for opportunities.
- The approach saved tens of thousands of user-hours of waiting per day.