One CLAUDE.md file. Keeps Claude responses terse. Reduces output verbosity on heavy workflows. Drop-in, no code changes.
5.8k
Stars
445
Forks
2
Open issues
6
Contributors
AI Analysis
A single-file prompt optimization tool (CLAUDE.md) that reduces token consumption by instructing Claude to generate terse, structured responses without filler. Designed specifically for automation pipelines and high-volume output workflows where the persistent instruction cost is offset by reduced output tokens. Best for teams running repeated Claude API calls at scale; not suitable for casual one-off queries or exploratory work.
Inferred from signals mentioned in the README (tests, CI, type safety) — not a review of the actual code.
AI's overall editorial judgment — not an average of the bars above, can weigh other factors too.
Single-file Claude prompt config cuts output verbosity for token-heavy automation workflows
claude-token-efficient provides a single CLAUDE.md instruction file that reduces Claude's default verbose response patterns — filler phrases, over-engineering, restatements — to lower output token costs in high-volume automation pipelines. It targets developers running agent loops, code-generation pipelines, or batch tasks where Claude's default verbosity compounds across hundreds of calls. The project is honest about its limits: net savings only materialize when output volume is high enough to offset the persistent input token cost of loading the file each turn. Real measured output-token reductions are modest (~4–12% on API models), not the 63% headline figure.
Created March 2026, roughly coinciding with a surge of Claude Code tooling. The project evolved from a simple rules snippet to include reproducible benchmarks and semantic evals, responding to community scrutiny in issues.
5,700+ stars in under three months signals strong organic interest, likely driven by developer frustration with LLM token costs and the appeal of a zero-code, drop-in solution. The timing aligns with the broader Claude Code tooling ecosystem expanding rapidly in early 2026. Recent star velocity (~61/week) suggests the initial spike has normalized but interest remains steady rather than collapsing.
An independent external benchmark (linked in README via Issue #1, attributed to github.com/adam-s/testing-claude-agent) tested cost-to-green across 3 coding challenges, providing some third-party validation. 436 forks suggest active experimentation by other developers. However, documented production deployments at scale are not verified beyond this. The benchmark comparison data exists but is not equivalent to confirmed production usage evidence.
Appears to be a minimal Python repository containing a CLAUDE.md instruction file, a benchmark runner script (benchmark/run.py), and associated summary reports. Likely no complex runtime logic — the core deliverable is the instruction file itself, not executable code. The Python label likely reflects the benchmark tooling rather than a substantial codebase.
Benchmark scripts exist (benchmark/run.py, SUMMARY.md, SEMANTIC.md) that provide reproducible API-level token measurements across models at N=5 runs. This is functional validation rather than unit test coverage. Traditional software test coverage is not documented in README.
Last push June 16, 2026 — 10 days before evaluation date. Active within the past two weeks. README shows iterative refinement: methodology notes were added, benchmarks were revised downward when community testing challenged the 63% claim, and semantic evals were added. These are signs of a maintainer responding to feedback rather than abandoning the project.
ADOPT IF: you run high-volume Claude automation pipelines (agent loops, batch code generation) with persistent sessions where output verbosity compounds across many calls, and you want a zero-code way to reduce that verbosity. AVOID IF: your use case involves short or infrequent sessions, exploratory/architectural conversations, or you need guaranteed parseable output at scale (use structured outputs instead). MONITOR IF: you are already using a similar CLAUDE.md or system prompt — review the semantic eval results to confirm which rules still affect current model behavior, since Anthropic model updates may render some rules obsolete.
Independent dimensions
Mainstream potential
3/10
Technical importance
4/10
Adoption evidence
3/10
- Model drift: Anthropic model updates may change default verbosity behavior, potentially making some rules redundant (the README's own semantic eval already shows current models have eliminated several targeted behaviors like sycophantic openers). The file's value erodes if the model already behaves tersely by default.
- Overstated headline benchmark: The 63% word-reduction figure was a single-run directional test, not a controlled study. Real API output-token reduction is 4–12% per reproducible benchmarks — users relying on the headline number may be disappointed in practice.
- Input token overhead: On low-volume or conversational use, the persistent input cost of loading CLAUDE.md each turn produces a net token increase, not savings. Misuse in inappropriate contexts costs money rather than saving it.
- Narrow scope: The project solves a real but narrow operational problem. As Claude Code and similar tools mature, official configuration or system-prompt management features from Anthropic may subsume this use case entirely.
- Benchmark reproducibility: Token counts vary between runs and model versions. The N=5 benchmark provides directional evidence but may not generalize across all prompt types, domains, or future model revisions.
Likely to remain a useful lightweight utility for Claude-heavy pipeline developers in the near term. Long-term relevance depends on whether Anthropic introduces first-party verbosity controls. Growth will probably plateau or slow as the Claude Code tooling space matures and similar configs proliferate.
Newsletter
Get analyses like this every Monday
Free weekly digest of the most interesting open-source discoveries.
Languages
Information
- Language
- Python
- License
- MIT
- Last updated
- 3w ago
- Created
- 3mo ago
- Analyzed with
- anthropic/claude-haiku-4-5
Stars over time
Contributors over time
Top 100 contributors only — repos with more will plateau at 100.
Open issues
Open pull requests
No open pull requests.
Top contributors
Recent releases
No releases published yet.
Similar repos
Maciek-roboblog/Claude-Code-Usage-Monitor
Claude Code Usage Monitor is a real-time terminal utility that tracks API...
multica-ai/andrej-karpathy-skills
A single CLAUDE.md configuration file designed to guide Claude Code...
trailofbits/claude-code-config
This repository provides opinionated defaults, documentation, and workflows for...
SuperClaude-Org/SuperClaude_Framework
SuperClaude is a configuration framework that augments Claude Code with 30...
| Repository | Stars | Week Δ | Language | Score | Updated |
|---|---|---|---|---|---|
|
drona23/claude-token-efficient
this
|
5.8k | +50 | Python | 7/10 | 3w ago |
|
Maciek-roboblog/Claude-Code-Usage-Monitor
|
8.4k | — | Python | 8/10 | 5d ago |
|
multica-ai/andrej-karpathy-skills
|
190.3k | — | — | 6/10 | 3mo ago |
|
trailofbits/claude-code-config
|
2k | — | Shell | 8/10 | 3mo ago |
|
SuperClaude-Org/SuperClaude_Framework
|
23.5k | — | Python | 7/10 | 4w ago |
 davila7/claude-code-templates
|
28.6k | — | Python | 7/10 | 8h ago |
28k stars; broader in scope — provides reusable prompt templates for various tasks rather than focusing on token efficiency. Complementary rather than directly competing.
2k stars; security-focused Claude config from a credible organization. More specialized around safe coding practices than cost reduction. Different primary goal.
8.9k stars; tips and techniques in HTML format rather than a drop-in config file. Educational resource vs. operational artifact — different use pattern.
8.3k stars; monitors and reports token usage rather than reducing it. Complementary tool — monitor can help verify whether this file's savings are real in a given workflow.
First-party API feature. README explicitly recommends this over CLAUDE.md for parser reliability at scale. More robust for guaranteed output format but requires code changes — this file is a no-code alternative.