Claude Code built this to explain prompt caching
I asked Claude Code to explain prompt caching. It built me this machine.
One line, with a free skill called Anatomy
One line, and a free skill called Anatomy.
First call: it reads every tile. $0.104
Each tile is a thousand tokens of the prompt. The first call reads every one, and pays a little extra to cache them: about ten cents.
Cached: $0.008 a call
Call again, and the clamped tiles are already cached. The reader skips them, and the call costs less than a cent.
The catch: the cache lasts 5 minutes
The catch: the cache only lasts five minutes. Wait any longer, the clamps let go, and the next call pays to read it all again.
22 minutes. One prompt.
Our run, 14 October
About 70 parts, on Opus 5.5. It checked every line of its own drawing before it stopped.
It took twenty-two minutes on its own. And it checked every line of its own drawing before it stopped.
npx skills add wheresryan22/anatomy
Free · MIT · Claude Code, Codex, OpenCode
About 9,900 GitHub stars in six days.
github.com/wheresryan22/anatomy ↗Anatomy is free, and it runs in Codex and OpenCode too. It's at almost ten thousand GitHub stars in under a week.
What would you have it explain?
Your turn
What would you have it explain? Tell me below.
