vet_source
Classify the licence: Tier 1/2/3 or Excluded. Paywalled / all-rights-reserved sources are refused and routed to a citation-only signpost.
Turn a vetted authoritative source (a standard, handbook, guidebook, or framework) into a licence-clean, citable knowledge pack your agent loads on demand.
A fork of book-to-skill, repositioned from personal study skills to publishable reference oracles: it vets the source's licence first and refuses to package anything it cannot redistribute.
A reference pack answers "what does this body of knowledge say about X?" from the actual source: grounded, never hallucinated, never a photocopy. It keeps book-to-skill's progressive-disclosure shape (a small always-loaded SKILL.md plus on-demand chapters) and adds the provenance, licence, and verification rigour that a published pack needs.
Classify the licence: Tier 1/2/3 or Excluded. Paywalled / all-rights-reserved sources are refused and routed to a citation-only signpost.
Vendored book-to-skill engine turns PDF/EPUB/DOCX/HTML/RTF into clean text + metadata.
Deterministic table of contents with exact char/line offsets, so chapters are sliced precisely, not by ad-hoc grep.
Vet-gated scaffold writes PACK.yaml provenance and a per-pack LICENSE carrying the source's terms.
Citation-grounded chapters plus glossary, patterns, cheatsheet, and the pack's own SKILL.md index.
Three gates: no verbatim overlap, valid structure + tier, and every topic-index route grounded in its chapter.
Tier 1/2/3/Excluded classification before extraction. An Excluded source is a hard stop: it becomes a signpost, never a pack.
PACK.yaml records title, publisher, version, licence, tier, and NC/SA flags; each pack ships the source's own LICENSE.
N-gram shingling flags any run lifted from the source: a licence-safety and quality gate, enforced, not just advised.
Every topic-index route is verified to be grounded in the chapter it points to, so the pack routes queries honestly.
Claude Code, OpenClaw, GitHub Copilot CLI: the skill folder (tools included) is copied unchanged.
Codex, Gemini CLI, Cursor: the spec is rendered into the agent's prompt/command format. Tools run from a local clone.
Built on the agentskills.io SKILL.md standard, so the same source installs everywhere via install.py.
# Clone, then install into your agent (Claude Code by default) git clone https://github.com/jgsystemsconsulting/jgs-reference-skill cd jgs-reference-skill python install.py # Claude Code python install.py --agent all # every user-global agent python install.py --list-agents # show each target + format
Restart your agent and run /jgs-reference-skill. Requires Python 3.9+. The tooling is MIT, but packs you generate carry their source's licence; read the vetting rubric before packaging anything.
Once installed, restart your agent and talk to the skill by name. Give it the source's title, publisher, and licence; it vets the licence first and refuses anything it cannot redistribute, routing Excluded sources to a citation-only signpost.
# load the generator spec /jgs-reference-skill # build a pack from a vetted source /jgs-reference-skill build a pack from the NASA SE Handbook # vet a source before committing to it /jgs-reference-skill vet "SysML" published by the Object Management Group # fold a new document into an existing pack /jgs-reference-skill add NIST SP 800-160 Vol 2 to my nist-sse pack
The skill runs the whole pipeline for you: vet, extract, outline, scaffold, generate, and the three verification gates. Prefer to run the tools by hand? Every step is a stdlib script with a --self-check. See the usage guide for the full reference.