Awesome DeepSeek Skills

中文


A small, pinned catalog of Agent Skills that can be loaded by DeepSeek Harness (DSH). Every entry has an immutable source commit, a license pointer, a static report, and basic network/command/write hints.

This is deliberately not a scrape of every repository containing SKILL.md. The first catalog is small enough to reproduce in CI.

Browse

See [the generated catalog](CATALOG.md), or list the machine-readable entries:

uv sync --locked
uv run python scripts/catalog.py list

The initial set includes DeepSeek protocol diagnostics and a few established design and development skills from Anthropic, Superpowers, and Trail of Bits.

Use it inside DSH

Replace demo with your profile:

dsh plugin --profile demo add github:Whning0513/awesome-deepseek-skills

After restarting DSH, two tools are available:

• deepseek_skills_list searches the bundled catalog without going online.
• deepseek_skill_install fetches one pinned source and installs it in the current project.

For example, ask DSH to list the development entries, then install the id you choose. The install tool only writes to .agents/skills or .dsh/skills; it refuses an existing destination and does not run anything from the skill. Git is required. Python is not required for the DSH plugin.

Install one pinned skill

The installer checks out the exact commit in the entry, refuses to overwrite an existing directory, rejects symlinked files, and runs the DSH profile before copying anything. It does not execute skill scripts.

uv run python scripts/catalog.py install \
  obra/superpowers/verification-before-completion \
  --target /path/to/project/.agents/skills

DSH discovers direct children of .agents/skills and .dsh/skills. The installer records provenance in .deepseek-skills.lock.json next to the installed bundles.

The Python installer and DSH plugin use the same catalog, pinning rules, and lock-file format. They are tested on Linux, Windows, and macOS.

What “verified” means here

Each pinned source is checked with:

• the Agent Skills reference parser used by deepseek-skill-doctor;
• skill-validator v1.6.0, structure mode only;
• the DeepSeek/DSH compatibility profile.

Structure mode is intentional: external link checks depend on live websites and are not reproducible. Warnings stay in the report and the catalog table. No model is used as a judge.

This is not a security audit. A skill can contain unsafe instructions or scripts even when its format is valid. Read the source, the risk hints, and the license before installing it.

Reproduce the reports

CI downloads the pinned upstream validator binary and verifies its published SHA-256. Locally, pass the same binary path:

uv run python scripts/catalog.py verify \
  --skill-validator /path/to/skill-validator \
  --verified-at 2026-08-14 \
  --check

Add a skill

See [CONTRIBUTING.md](CONTRIBUTING.md). A submission needs a public source, immutable commit, license evidence, a useful DSH trigger in description, and a reproduced report. New projects are welcome, but they stay marked new; age and popularity are not invented by the catalog.

MIT licensed. Catalog entries retain their own licenses. This is a community project, not an official DeepSeek component.