Community Skill Harvesting

Community Skill Harvesting

Trigger: Use when importing existing SKILL.md files from public GitHub skill repos into your personal or organization skills repository.

When to Use

• Expanding your skills library by sourcing from known community repos
• Curating the best skills from mattpocock/skills, anthropics/skills, emilkowalski/skills, MiniMax-AI/skills, slavingia/skills, google/skills
• Creating a merged, deduplicated repository from multiple upstream sources
• Setting up a one-time bulk import from a new upstream repo

Not for single skill installs via hermes skills install <url>.

Key Sources

Non-harvestable: vercel-labs/skills (CLI tool), openai/skills (deprecated), awesome-openclaw-skills (index only).

Recently discovered harvestable:

Procedure

1. Survey

curl -s "https://api.github.com/repos/owner/repo/contents/skills" | \
  python3 -c "import sys,json; [print(d['name']) for d in json.load(sys.stdin) if d['type']=='dir']"

2. Build Import Map

SOURCES = {
    "source/repo": {
        "category": ["skill1", "skill2"],
    },
}

3. Fetch and Enhance

from urllib.request import urlopen

def fetch_skill(repo, path, name):
    url = f"https://raw.githubusercontent.com/{repo}/main/skills/{path}/SKILL.md"
    with urlopen(url, timeout=15) as r:
        return r.read().decode("utf-8")

enhanced = f"""---
name: {name}
description: {desc}
source: {repo}
tags: [engineering]
metadata:
  hermes:
    tags: [engineering, design-patterns]
---

{body}
"""

4. Handle Name Collisions

NAME_ALIASES = {
    "prototype": {"mattpocock/skills": "prototype-solution"},
}
mapped = NAME_ALIASES.get(name, {}).get(repo, name)

5. Fetch Supporting Files

for sub in ["references", "templates", "scripts", "assets", "examples"]:
    url = f"https://api.github.com/repos/{repo}/contents/skills/{path}/{sub}"
    try:
        with urlopen(url, timeout=10) as r:
            for e in json.loads(r.read()):
                if e["type"] == "file" and e["name"] != "SKILL.md":
                    # download e["download_url"] to target/sub/e["name"]
                    pass
    except:
        pass

6. Process in Parallel (for 50+ skills)

For large hauls (50+ skills across multiple repos), dispatch parallel subagents:

from hermes_tools import delegate_task

# Split skills across 4 agents (~25 skills each)
batches = [
    skills[0:25], skills[25:50],
    skills[50:75], skills[75:100],
]
for batch in batches:
    delegate_task(
        goal=f"Read, enhance, and install {len(batch)} skills into Hermes Agent...",
        role="leaf",
    )
# Results deliver back when all complete — do not poll.

Each subagent gets isolated context and terminal session. The final summaries arrive as messages in the conversation.

7. Verify Counts

ls skills/ | wc -l
# Confirm expected total: existing + imported = final count

python scripts/import-community-skills.py
git add -A && git commit -m "Import N skills from {sources}" && git push

Pitfalls

• GitHub rate limits: 60/hr unauthenticated; 5000/hr with GITHUB_TOKEN
• Repo clone naming conflicts: Multiple repos may all be named skills/. Always clone with a named target dir: git clone --depth 1 <url> <repo-name>-skills
• MSYS2/git-bash path persistence: In git-bash, cd "$WORKDIR" changes don't persist between command invocations. Always use explicit absolute paths (/c/Users/...) rather than relative cd state.
• False-positive skill detection: A SKILL.md filename doesn't guarantee real skill content. Filter by: has frontmatter (---), has substantive body (>200 chars), has sections (## ), has meaningful description. Reject non-skill files like CHANGELOG.md, CONTEXT.md, templates, LICENSE files.
• Repo structure varies: Flat vs categorized — adjust paths per source
• License compliance: Preserve licenses and add source: attribution
• Name collisions: Check all names before bulk import
• Stale content: Re-check upstream quarterly
• Branch naming: Some repos use master not main
• Idempotency of frontmatter enhancement: Re-running enhance_frontmatter on already-enhanced content duplicates metadata: blocks if the entire old metadata subtree isn't dropped — orphaned nested keys (e.g. related_skills: under hermes:) survive the filter and a new metadata: block gets appended on every run. Always drop the complete indented subtree under metadata: (every following line indented more than metadata: itself), not just lines matching hermes:/tags:/category: by name. Verify by round-tripping: enhance(content) == enhance(enhance(content)).

Verification

ls skills/ | wc -l
grep -l "^source:" skills/*/SKILL.md | wc -l
python3 -c "from pathlib import Path; n=[d.name for d in Path('skills').iterdir() if d.is_dir()]; print([x for x in set(n) if n.count(x)>1] or 'OK')"

Supporting Files

• references/enhancement-pipeline.md — Adding Hermes format (trigger, procedure, pitfalls, verification) to imported skills. Documents known bugs: body changes lost when FM updated, false-positive "How to" detection, and proper section matching patterns.
• references/source-definitions.md — Full source map and path mappings for all harvested repos.
• references/resume-from-local-clones.md — When GitHub raw fetches are flaky or interrupted, resume the import from a local staging clone directory instead of re-fetching online. Iterate rglob("SKILL.md") over each cloned repo dir, apply enhance_frontmatter, and verify idempotency via round-trip equality before writing. The staging dir acts as a checkpoint: if the import is interrupted midway, re-running picks up exactly where it left off (existing target files are skipped).
• scripts/import_community_skills.py — The bulk import script. Run from repo root.
• scripts/resume_skill_import.py — Resumable import script that reads from local clones (no network fetches), applies frontmatter enhancement idempotently, and copies supporting files (references/templates/scripts/assets/examples).

Related Skills

• skills-repo-setup — Structural repo setup for GitHub
• skill-factory-system — Batch skill creation from trend analysis
• hermes-agent-skill-authoring — Authoring individual SKILL.md files