GitHub's drona23/claude-token-efficient offers a simple solution to slash Claude's verbosity. Just drop one file, `CLAUDE.md`, into your project root and output tokens drop by roughly 63% across common prompts. The file targets Claude's default habits: sycophantic openings, redundant closings, em dashes, and over-engineered code suggestions. It's a drop-in fix requiring no code changes, though it consumes input tokens on every message. The biggest savings appear in high-output scenarios like automation pipelines and code reviews. Benchmarks show a 63% reduction in word count for key tasks like explaining async/await or code reviews, though results vary by prompt and usage context. The project's creator emphasizes it's best for repeated structured tasks, not single queries or exploratory work where Claude's default verbosity is intentional.

The file's impact is most tangible in high-volume workflows. For example, a code review prompt dropped from 120 words to 30 words with `CLAUDE.md`, saving significant tokens. However, the persistent input cost means low-output usage becomes net negative. The repository includes profiles tailored to different needs - `CLAUDE.coding.md` for developers, `CLAUDE.agents.md` for automation - and allows rules to be layered globally or project-specific. Community contributions are encouraged, with specific rules targeting actual pain points rather than generic conciseness.

Practical application requires careful consideration. While the 63% reduction is a compelling directional signal, it's not a universal guarantee. The tool excels at fixing predictable failures like off-by-one errors or redundant explanations but won't resolve deeper issues like hallucinations or architectural drift. Users should deploy it where Claude's default output is genuinely wasteful and where the input cost is offset by substantial output savings. The project provides clear benchmarks and usage guidelines, making it a valuable resource for token-conscious Claude users, particularly in development and automation contexts.