anbeime/skill Is a Scraper, a Local Skill Dump and a Static Index in One Repository

The problem it solves: skill discovery across dozens of upstream teams

Agent skills are scattered. Anthropic publishes docx, pptx, xlsx, pdf and mcp-builder skills. Vercel publishes react-best-practices and composition-patterns. Cloudflare publishes agents-sdk and durable-objects. Trail of Bits publishes building-secure-contracts and static-analysis. The README lists 182 official skills pulled from the awesome-agent-skills repository, plus 63 local skills, for a stated total of 245. The value here is not that the repository wrote those skills. It did not. The value is that one Python program walks the upstream list on a schedule and writes the result into data/skills.json, so you can grep one file instead of opening a dozen vendor repositories. The intended audience is a developer who already runs a skill-aware agent and wants a local catalogue with categories, source attribution and an update timestamp. It is not aimed at someone who wants a package manager, because there is no install step for individual skills.

What the crawler actually does, and what it does not do

The mechanism described in the README is narrow: every 24 hours the project fetches the awesome-agent-skills repository and records what it finds. The command surface is main.py with a single flag per mode. There is no plugin API, no webhook, and no incremental diff described in the material. The README claims a 100% backup coverage figure backed by 71 archives, but those archives live at D:\tool\skills\skill压缩文件\, a Windows path that will not exist on a Linux CI runner or a macOS laptop. That is the first real constraint: the backup story is a personal filesystem convention, not a portable feature. The second is that the local skill count is broken down as 27 core plus 30 sub-skills plus 6 system built-ins, which sums to 63. That arithmetic is consistent, but it also tells you the local library is a curated personal collection rather than a community-contributed one. Nothing in the material describes a contribution process for adding a skill to the local set.

Getting it running: clone, install, run once or as a daemon

The README gives a four-step quick start. Clone with git clone https://github.com/anbeime/skill.git, then cd skill. Install with pip install -r requirements.txt. Then run one of five modes: python main.py --once performs a single update; python main.py --daemon starts the scheduled update process; python main.py --stats prints data statistics; python main.py --export skills.csv writes a CSV; python main.py --once -v turns on verbose logging. Browsing is file-based: cat data/skills.json for the crawled set, cat data/local_skills.json for the curated set, and ls docs/ for the documentation directory. The README also documents a validator: python tools/skill_validator.py validate checks every skill, and the same tool accepts a single skill argument, though the README text is truncated mid-command at python tools/skill_va. Note the mismatch between the badge and the quick start: the badge advertises an online store at skill.vercel.app while the repository homepage field points at skill.miyucaicai.cn. Both are mentioned, neither is described further.

The licences of the 245 skills are the weakest part of the pitch

The repository itself shows a CC-BY-4.0 badge in the README, but the licence field for the repository is unknown and no LICENSE file is confirmed in the material. That matters more here than in a normal project, because this repository redistributes other people's skills. Anthropic, Vercel, Cloudflare, Stripe, Supabase, Expo, Sentry, Hugging Face and Trail of Bits each publish under their own terms, and a CC-BY-4.0 badge on the aggregator does not relicense the aggregated content. The README also lists a 知易智能基座 product with a Feishu application link, which suggests the repository doubles as a funnel for a commercial service. That is not disqualifying, but it means the repository is not a neutral index. If you plan to ship any of the local skills inside a product, the material does not give you a per-skill licence field to check. Treat the aggregated set as a discovery index you read, not a bundle you redistribute, until a LICENSE file and per-skill attribution appear.

Where the local skill packs sit, and why the star ratings are not evidence

The 63 local skills are grouped into roughly a dozen categories. Content creation has ten entries including content-creation-publisher and baoyu-post-to-wechat. Video creation has nine, including video-creation-suite and three-body-video-creator. E-commerce has seven, headlined by ecommerce-full-pipeline, which the README describes as covering 1688 collection, cleaning, listing, promotion, video, dropshipping and Xianyu product picking. Document and analysis has four, including paper-analysis-assistant for arXiv papers and contract-review. Software development has only two: ontoly-software-graph for architecture review and request tracing, and archify, which converts a codebase into interactive architecture, workflow, sequence and data-flow diagrams exported as single-file HTML plus PNG, SVG and WebM. Every entry carries a star rating from two to five. Those ratings are the README author's own labels, not a measurement, and the repository does not document who assigns them or against what criteria. Read them as a hint about which packs the maintainer uses most, nothing more.

The alternative is the upstream list, and the difference is aggregation versus curation

The obvious alternative is awesome-agent-skills itself, which this project crawls every 24 hours. The difference in approach is straightforward. Upstream is a human-maintained list of links, and it is the source of truth for the 182 official skills. anbeime/skill adds three things on top: a machine-readable export in JSON and CSV, a classification layer by function, source and star count, and a local set of 63 Chinese-language skill packs that upstream does not carry at all. If you only need the official Anthropic and Vercel skills, going upstream gets you there with fewer moving parts and no Python dependency. If you need the Chinese content, video and e-commerce packs, or you want to feed a CSV of skills into your own tooling, upstream will not help you. The trade is that you inherit this project's 24-hour lag and its parsing assumptions about upstream's format, neither of which is documented.

Maintenance cost is low, but the update path is a single maintainer's filesystem

Running this is cheap. One Python process, one scheduled flag, one JSON output. The costs that matter are elsewhere. First, the crawler breaks when upstream restructures, and the material offers no schema, no version pin and no error-handling description for that case. Second, the backup feature assumes D:\tool\skills\ exists, so on any non-Windows host the 100% coverage claim is inapplicable. Third, the documentation set is written in Chinese (技能管理数据库.md, 技能清理与迁移指南.md, 技能数量差异分析报告.md, specification.md, best-practices.md, quickstart.md), which is fine for the target audience and a barrier for anyone else. Fourth, there are no releases retrieved, so upgrades mean pulling main. There is no changelog in the material and no stated compatibility policy for the JSON shape that data/skills.json exposes. If you build on that file, snapshot it yourself.

Editorial conclusion

Adopt anbeime/skill if you want a single local index of agent skills that refreshes on a timer and can be exported to CSV for your own tooling. Do not adopt it if you need a per-skill licence you can audit before shipping, since the repository carries no LICENSE file and only a CC-BY-4.0 badge in the README. Before cloning, verify three things: whether a LICENSE file now exists at the repository root, whether requirements.txt and main.py both exist on main, and whether tools/skill_validator.py accepts your own SKILL.md, because that validator is the only quality gate this project offers.

Official sources

1. anbeime/skill on GitHub
2. Issues
3. Project website
4. README