claude-token-efficient

claude-token-efficient is a repository built around one file: a CLAUDE.md you drop into a project to keep Claude's responses terse. Its README says this can reduce total tokens on output-heavy workflows. The problem it describes is that, by default, every word Claude generates in Claude Code costs tokens, and most people never control how Claude responds.

There are two ways to apply it. Option 1 pastes a short rules block into a chat session for one-off tasks, for example read files first, write a complete solution, test once, no over-engineering. Option 2 places the CLAUDE.md file in the project root for a set-and-forget setup.

The README is open about limits. It warns that instruction files add input tokens on every turn and that a long file can cost more than it saves. Benchmarks were run on Claude only, and results on other models are untested. The original table is described as a 5-prompt directional indicator, not a statistically controlled study, and a later reproducible benchmark folder reports output tokens across Claude models.