An open-source project called claude-token-efficient claims to cut Claude API coding costs by up to 17% by stripping conversational filler and reducing unnecessary code iterations, according to benchmark data published in the project's own GitHub repository. The repository, created by developer drona23, reports that its optimized CLAUDE.md configuration reduced spend from $1.131 to $0.935 across three coding tasks — a CSV reporter, SQLite window functions, and a WebSocket counter. These are the project's self-reported figures, not an independently verified benchmark.
Eliminating Conversational Fluff and Over-Engineering
Default model responses often include polite preambles, concluding remarks, and unrequested abstractions that consume output tokens. The repository provides a lightweight CLAUDE.md system-prompt file intended to enforce concise behavior across project workflows. According to the repo's documentation, it targets:
- Sycophantic chatter: removing openers like "Great question!" and closers like "I hope this helps!"
- Unnecessary file rewrites: preferring targeted edits over full-file rewrites
- Over-engineered solutions: favoring simple, direct fixes over unrequested complexity
- Unvalidated completion: requiring tests to run before a task is marked done, to avoid costly debugging loops
Benchmark Results and Trade-offs
According to the repository's published comparison of six prompt configurations, the M-drona23-v8 configuration had the lowest total cost, saving a reported 17.4% versus rival prompt structures, mainly by reducing iterative debugging cycles.
The repo also notes a trade-off: instruction files add input tokens to every turn, since the rules load into context each message. For short or low-volume queries, this could produce a net cost increase rather than a saving — the benefit reportedly only appears when output volume is high enough to offset the added input overhead.
The project also reports that pasting the same rules directly into the chat prompt, rather than using a persistent CLAUDE.md file, increased total task cost by 41% ($1.318 vs. $0.935), which it attributes to losing out on context-caching efficiencies.
Implementation
The repository suggests placing a minimal CLAUDE.md file in a project root or in a global directory (~/.claude/CLAUDE.md), with rules composing across levels so global preferences can stay separate from task-specific constraints. It also suggests a simple inline directive for quick workflows: Rules: Read files first. Write complete solution. Test once. No over-engineering. The documentation cautions that prompt-level rules are not a substitute for native structured outputs or API schemas when strict JSON parsing is required.