zhangdw/astra-skills

🧰 ASTRA Skills: Agent Skill Tool-use Repository Atlas

A large-scale atlas of real-world AI agent skills discovered from public GitHub repositories

    

ASTRA Skills (Agent Skill Tool-use Repository Atlas) is a deduplicated corpus of agent skill directories centered on SKILL.md, collected with the Astra crawler for research on tool use, agent instruction design, and skill retrieval.

Quick Start · At a Glance · Directory Format · Intended Use

---

This dataset packages skill directories discovered from public GitHub repositories. The dataset-level metadata and packaging are released under Apache-2.0, but individual skill contents remain subject to their original repository licenses and terms. Please inspect upstream licenses before redistribution, training, or benchmark release.

✨ Why ASTRA Skills?

Modern coding agents increasingly rely on reusable skills: small instruction bundles, scripts, templates, and examples that teach an agent how to perform a specialized workflow. These skills are scattered across repositories and marketplaces, making large-scale study difficult.

ASTRA Skills provides a compact snapshot for questions like:

• What do real-world agent skills look like across public repositories?
• How are skills structured around SKILL.md, scripts, templates, and examples?
• Which instruction patterns, tool-use conventions, and retrieval signals appear in the wild?
• How can skill discovery, ranking, routing, composition, or deduplication methods be evaluated at scale?

📦 Dataset at a Glance

148,134 deduplicated skills	19 split archive parts	3.65 GiB compressed archive	GitHub source label
2026-04-15 snapshot date	09:10:28 → 11:59:38 UTC DB crawl range	MD5 SKILL.md dedup key	Directory Corpus skills plus metadata

Snapshot Summary

Field	Value
Total deduplicated skills in skills.db	148,134
Source labels in DB	github only
Crawl time range, UTC	2026-04-15 09:10:28 → 2026-04-15 11:59:38
Archive layout	astra-skills-part-0001.tar.gz → astra-skills-part-0019.tar.gz

🗂️ Dataset Files

File	Description
README.md	Dataset card and usage notes.
astra-skills-part-0001.tar.gz → astra-skills-part-0019.tar.gz	Split compressed archive containing the github/ skill directory tree.
.gitattributes	Hugging Face / Git LFS tracking metadata.

🚀 Quick Start

Download the dataset with the current Hugging Face Hub CLI:

uvx --from huggingface_hub hf download zhangdw/astra-skills \
  --type dataset \
  --local-dir astra-skills

Merge the split archive and extract it:

cat astra-skills/astra-skills-part-*.tar.gz > skills_github.tar.gz
tar xzf skills_github.tar.gz

This produces a github/ directory containing the collected skill directories.

Inspect a few skills:

find github -name SKILL.md | head
find github -name _meta.json | head

🧱 Directory Format

Each saved skill directory is copied from a GitHub repository path that contains SKILL.md.

github/
└── {source}_{owner}_{skill_name}/
    ├── SKILL.md
    ├── _meta.json
    └── [optional files, e.g. scripts/, templates/, examples/]

_meta.json stores crawl metadata. Typical fields include source site, repository URL, owner, repository name, skill name, and relative path inside the original repository.

🔎 Example Discovery Workflows

Build a lightweight local skill index

from pathlib import Path

root = Path("github")
records = []

for skill_file in root.rglob("SKILL.md"):
    records.append({
        "skill_dir": str(skill_file.parent),
        "skill_name": skill_file.parent.name,
        "readme_chars": len(skill_file.read_text(errors="ignore")),
        "has_scripts": (skill_file.parent / "scripts").exists(),
        "has_templates": (skill_file.parent / "templates").exists(),
    })

print(len(records))
print(records[:3])

Search skills by instruction text

rg -n "browser automation|Hugging Face|spreadsheet|PowerPoint" github -g 'SKILL.md'

Read upstream provenance from metadata

import json
from pathlib import Path

for meta_file in Path("github").rglob("_meta.json"):
    meta = json.loads(meta_file.read_text())
    print(meta_file.parent.name, meta)
    break

⚙️ Crawl Pipeline

The current snapshot was produced by the Astra crawler with the following high-level pipeline:

1. Discover GitHub repositories from skill index websites (skills.sh, skillstore.io, agent-skills.md) plus configured seed repositories.
2. Clone discovered repositories.
3. Recursively detect directories containing SKILL.md.
4. Copy each detected skill directory into the dataset layout.
5. Deduplicate by MD5 hash of SKILL.md content in skills.db.

Website index coverage does not guarantee successful retrieval of every listed skill. Repository accessibility, branch availability, file layout, and upstream content changes can all affect crawl completeness.

✅ Intended Use

ASTRA Skills is designed for:

• research on AI agent skill ecosystems and instruction-following behavior;
• skill retrieval, ranking, routing, and composition experiments;
• analysis of real-world tool-use instructions, scripts, templates, and examples;
• deduplication, clustering, and taxonomy construction for agent skills;
• benchmark construction for skill discovery and agent memory systems.

🛠️ Maintenance Notes

This dataset is a snapshot, not a live mirror. The archive preserves the collected directory contents from the crawl, while _meta.json files preserve repository-level provenance. If you build downstream datasets from ASTRA Skills, keep the snapshot date and upstream license caveats with your derived artifacts.

👤 Author

• Dawei Zhang (GitHub: zhangdw156)

📚 Citation

If you use ASTRA Skills in research, please cite this dataset and any upstream repositories whose skill contents are central to your analysis.

@misc{astraSkills2026,
  author       = {Dawei Zhang},
  title        = {ASTRA Skills: Agent Skill Tool-use Repository Atlas},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/8399},
  url          = {https://huggingface.co/datasets/zhangdw/astra-skills},
  note         = {Snapshot date: 2026-04-15; DOI record revision: 146eb8c}
}

📄 License

Dataset metadata and packaging are released under Apache-2.0. Individual skill contents remain subject to their original repository licenses.

---

ASTRA Skills aims to make agent skill ecosystems easier to inspect, retrieve, and study at scale.