Anthropic says it made claude.ai 3x faster in two weeks
The write-up includes the prompts and methods, which is the part worth testing on your own codebase.
Anthropic's ClaudeDevs account says it made claude.ai 3x faster in two weeks, and published how the team used Claude itself to measure, debug and improve performance. The post includes the prompts and methods.
Boris Cherny framed it for people outside Anthropic, pointing at both the web app and the Desktop app. "If you've noticed how fast https://t.co/9A6X5iqpbs and the Desktop app have become in the last few weeks, here's how we did it," he wrote, describing the blog post as aimed at engineers working on speeding up their own apps.
Why the format matters more than the number
Thariq said the post is a deliberate experiment in a different kind of vendor write-up, one that shares specific prompts and techniques readers should be able to replicate, and asked for feedback on whether it is useful. That is the unusual part. A 3x number on a vendor's own product is a claim nobody outside the company can check, but a prompt is something you can paste into your own repo this afternoon and watch fail or work.
For working engineers, the thing to watch is whether performance work generalises the way the post implies. Measuring and debugging latency depends heavily on the stack, the tooling and what profiling data an agent can actually see. Anthropic's team had the advantage of knowing their own codebase and having their model tuned on the kind of work they do. Nobody has yet reported trying the published prompts on an unrelated application, so there is no independent result to point at.
The broader signal is a shift in what a lab publishes. Most vendor posts stop at the headline result and leave the method vague enough that it cannot be reproduced or contradicted. This one hands over the inputs, which makes it falsifiable in a way marketing usually is not. If the prompts turn out to be generic advice dressed up as methodology, that will become obvious quickly. If they are not, it is a reusable recipe for a category of work, performance debugging, where agents have been less obviously useful than in feature work.
this is a new type of post we're trying where we share in-depth about how we do work using specific prompts and techniques you should be able to replicate
Lots of juicy learnings & techniques in the blog post for engineers working on speeding up your own apps.