zhangdw/astra-skills-scripts

🛠️ ASTRA Skills Scripts: Script-Backed Skills from the Agent Skill Tool-use Repository Atlas

Script-backed skills from the Agent Skill Tool-use Repository Atlas

    

ASTRA Skills Scripts contains only skill directories from zhangdw/astra-skills that include a scripts/ subdirectory, making it easier to study agent skills that pair written instructions with runnable helper code.

Quick Start · At a Glance · Subset Definition · Directory Format

---

This dataset is a filtered subset of the public ASTRA Skills Collection. The dataset-level metadata and packaging are released under Apache-2.0, but individual skill files and scripts remain subject to their original repository licenses. Inspect upstream licenses before redistribution, execution, training, or benchmark release.

✨ Why a Scripts Subset?

Many agent skills are pure instruction documents. Others include executable helpers, shell scripts, Python utilities, templates, or workflow code under scripts/. The scripts-backed subset is useful when the research question is not only what agents are told to do, but how skills operationalize those instructions through local tools.

This subset helps answer questions like:

• Which real-world skills rely on runnable helper scripts?
• How do SKILL.md instructions describe when and how to call scripts?
• What tool-use patterns appear around code generation, browser work, document editing, data processing, and automation?
• How can retrieval systems surface skills that include executable affordances, not just prose instructions?

📦 Dataset at a Glance

28,954 skills with scripts/	4 split archive parts	1.84 GiB compressed archive	Strict Subset of zhangdw/astra-skills
19.55% of the full corpus	2026-04-15 extraction date	GitHub source lineage	Script-Backed agent workflows

Snapshot Summary

Field	Value
Source dataset	zhangdw/astra-skills
Full corpus size	148,134 deduplicated skills
Matched skills with scripts/	28,954
Subset share	19.55% of the full corpus
Extraction date	2026-04-15
Source root scanned	astra-skills/github/
Copied subset root	github/
Archive layout	astra-skills-scripts-part-0000.tar.gz → astra-skills-scripts-part-0003.tar.gz

🧩 Subset Definition

Rule	Description
Parent dataset	The full zhangdw/astra-skills corpus.
Selection predicate	Keep a skill directory only if it contains a scripts/ subdirectory.
Directory contents	Preserve the skill directory, including SKILL.md, _meta.json, scripts/, and other copied files from the parent corpus.
Scope	A scripts-focused research subset, not a new crawl or live mirror.

🗂️ Dataset Files

File	Description
README.md	Dataset card and usage notes.
astra-skills-scripts-part-0000.tar.gz → astra-skills-scripts-part-0003.tar.gz	Split compressed archive containing the filtered 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-scripts \
  --type dataset \
  --local-dir astra-skills-scripts

Merge the split archive and extract it:

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

This produces a github/ directory containing script-backed skill directories.

Inspect a few script-backed skills:

find github -path '*/scripts' -type d | head
find github -name SKILL.md | head

🧱 Directory Format

Each saved skill directory comes from the parent ASTRA Skills corpus and includes a scripts/ subdirectory.

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

_meta.json stores provenance metadata from the original crawl, such as source site, repository URL, owner, repository name, skill name, and relative path inside the upstream repository.

🔎 Example Discovery Workflows

List common script file extensions

from collections import Counter
from pathlib import Path

counts = Counter()

for scripts_dir in Path("github").rglob("scripts"):
    if scripts_dir.is_dir():
        for path in scripts_dir.rglob("*"):
            if path.is_file():
                counts[path.suffix.lower() or "[no extension]"] += 1

print(counts.most_common(20))

Find skills whose instructions mention script execution

rg -n "scripts/|run .*script|execute|python|bash|node" github -g 'SKILL.md'

Build a lightweight metadata table

import json
from pathlib import Path

rows = []

for meta_file in Path("github").rglob("_meta.json"):
    skill_dir = meta_file.parent
    meta = json.loads(meta_file.read_text())
    rows.append({
        "skill_dir": str(skill_dir),
        "script_files": sum(1 for p in (skill_dir / "scripts").rglob("*") if p.is_file()),
        "meta": meta,
    })

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

✅ Intended Use

ASTRA Skills Scripts is designed for:

• research on script-backed agent workflows and executable skill affordances;
• skill retrieval and routing experiments that need a has_scripts signal;
• analysis of how SKILL.md instructions coordinate with helper code;
• studying tool invocation patterns, script conventions, and repository provenance;
• constructing smaller benchmark or training subsets from the full ASTRA Skills corpus.

🛠️ Maintenance Notes

This dataset is a static extraction from the full ASTRA Skills snapshot. It preserves the parent corpus layout while filtering for directories that include scripts/. Because scripts can execute arbitrary upstream code, treat this dataset as research material: inspect code before running it, sandbox execution, and retain upstream provenance when deriving new artifacts.

👤 Author

• Dawei Zhang (GitHub: zhangdw156)

📚 Citation

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

@misc{astraSkillsScripts2026,
  author       = {Zhang, Dawei},
  title        = {ASTRA Skills Scripts: Script-Backed Skills from the Agent Skill Tool-use Repository Atlas},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/8428},
  url          = {https://huggingface.co/datasets/zhangdw/astra-skills-scripts},
  note         = {Scripts-focused subset of ASTRA Skills; snapshot date: 2026-04-15; DOI record revision: 77d9d29}
}

For completeness, also cite the parent collection:

@misc{astraSkills2026,
  author       = {Zhang, Dawei},
  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 and scripts remain subject to their original repository licenses.

---

ASTRA Skills Scripts makes script-backed agent skills easier to inspect, compare, and study at scale.