Agent Skill Apprentice

A Windows-friendly workflow for evaluating GitHub repositories and turning high-signal automation patterns into reusable agent skills.

The project combines a skill package with a Python evaluation script. It can fetch GitHub metadata, prescreen candidate repositories, ask a local Claude CLI for a structured learning decision, and generate draft SKILL.md packages for the best candidates.

Why this exists

Agent ecosystems move quickly. Instead of bookmarking every interesting repository, this tool keeps a repeatable learning loop:

• find candidate repos from GitHub search reports or direct URLs
• score whether a repo contains reusable workflow knowledge
• produce a structured learn_now, watch, or skip decision
• scaffold a draft skill only when the repo clears the quality bar
• keep registry records for later audit

Quick start

python -m venv .venv
. .venv/Scripts/activate
pip install -r requirements.txt
pytest

To evaluate a repo with your local Claude CLI:

$env:CLAUDE_COMMAND="claude"
python skills/github-skill-apprentice/scripts/review_repo_for_skill.py --repo owner/repo --output-dir skill-runs

If your Claude executable is not on PATH, set CLAUDE_COMMAND to its full path.

Batch mode

python skills/github-skill-apprentice/scripts/review_repo_for_skill.py `
  --batch-report-json C:\reports\github_search_results.json `
  --batch-limit 20 `
  --shortlist-limit 5 `
  --output-dir skill-runs

The batch funnel keeps model calls focused by prescreening noisy repos locally.

Repository layout

skills/github-skill-apprentice/
  SKILL.md
  agents/openai.yaml
  references/output-schema.md
  scripts/review_repo_for_skill.py
tests/

Privacy and portability

This public version excludes historical run outputs, third-party repository snapshots, local registries, prompts, raw model responses, and machine-specific paths. Runtime credentials are read from environment variables only.