Vibeskills Benchmark — Development skill for Claude Code
Native vs. Lean Vibe on SkillsBench: paired reward results for 82 tasks and token/tool-call analysis for 81 tasks.
How to install Vibeskills Benchmark
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open foryourhealth111-pixel/vibeskills-benchmark and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Vibeskills Benchmark does
Native vs. Lean Vibe on SkillsBench: paired reward results for 82 tasks and token/tool-call analysis for 81 tasks.
Alternatives in Development
- Claude Code SDK TS — Configure models, enable tools, stream events, then fetch text, JSON, run details or token stats in one call v 206 ★
- Gemini CLI Integration Skill — This skill enables Claude Code to effectively orchestrate Gemini CLI (v0.16.0+) with Gemini 3 Pro for code gen 174 ★
- Caveman Micro — 6-line caveman micro prompt (85 tokens) that outperformed the original 552-token skill 162 ★
README
VibeSkills Benchmark: Native vs. Lean Vibe
This repository reports a paired comparison of Native and Lean Vibe on SkillsBench. Both conditions used the same global catalog of 195 skills. The reward analysis covers 82 tasks with a valid numeric verifier reward in both conditions, and the resource analysis covers 81 of those tasks with complete token and tool-call records.
Results
| Measure | Native | Lean Vibe | Difference |
|---|---|---|---|
| Mean verifier reward | 0.503 | 0.714 | +0.211 |
| Full-score tasks | 39/82 | 57/82 | +18 tasks |
| Total tokens, 81 tasks | 491.122M | 345.756M | -29.60% |
| Tool calls, 81 tasks | 9,954 | 6,664 | -33.05% |
Read the [paper, public data, figures, and reproduction instructions](studies/full-skills-comparison/README.md), or open the [Chinese study summary](studies/full-skills-comparison/README.zh-CN.md).
Experimental setup
SkillsBench normally provides a task-specific skill directory. For this study, the task-local skills were merged into one shared catalog of 195 unique skill names. This setup represents an environment in which a user has many installed skills and the agent must identify the relevant ones for each task.
Native used the standard task-execution path with access to the shared catalog. Lean Vibe used the same path and catalog, with one additional coach skill for goal clarification, task decomposition, skill selection, ordered execution, and deliverable checks. The task, verifier, execution framework, and underlying model were held constant.
Study materials
| Path | Contents |
|---|---|
| `paper.pdf` | Full paper |
| `paper.tex` | LaTeX source |
| `data/` | Public task-level data and data dictionary |
| `figures/` | Publication figures in PDF and PNG formats |
[scripts/reproduce.py](studies/full-skills-comparis |
Related Skills
Warden Bench
Run the golden benchmark suite for one token-warden agent (or all) and compare results against the frozen run1
Codex Token Meter
Native macOS menu bar apps for tracking Codex and Claude Code token usage, quota, cache efficiency, cost, and
Nse Trading Skills
9 Claude Code skills for Indian stock trading — technical analysis, position sizing, stop-loss, trailing stops
Mera Nse Trading Skills
9 Claude Code skills for Indian stock trading — technical analysis, position sizing, stop-loss, trailing stops
Pi Remote iOS
Native iOS companion and Pi extension for controlling an active Pi coding-agent session over a paired WebSocke
Hatch3r Benchmark
Run and analyze performance benchmarks. Compare results against baselines, identify regressions, and produce p
Related Agents
Bench Reporter
Updates benchmark documentation with results from bench-runner. Maintains BENCHMARK.md, README.md, and EVAL-GA
Skill Eval Reporter
Compares repeated paired execution results using blind A/B methodology and generates a skill effectiveness rep
Eval Auditor
Audits an evaluation setup (benchmark, A/B test, or model comparison) for methodology errors that would invali
