I built a tool that turns repeated file reads into 13-token references. My Codex and Claude Code sessions use 86% fewer tokens on file-heavy tasks.
I got tired of watching Claude Code re-read the same files over and over. A 2,000-token file read 5 times = 10,000 tokens gone.