Installable knowledge-pack skills for coding agents in the Nuclear domain, reconstructed from genuinely open sources, vetted, attributed, and shareable by construction. Ask a topic; your agent reads the right chapter and answers from reference notes, not a hazy recollection.
You are installing jgs-nuclear-knowledge-packs, an open catalogue (MIT tooling) of nuclear engineering knowledge-pack skills by JG Systems Consulting Ltd. Repository: https://github.com/jgsystemsconsulting/jgs-nuclear-knowledge-packs (version 0.3.0). Do this in order: 1. Read README.md, docs/skill-usage.md, and CHANGELOG.md so you understand what you are installing. There are NO external prerequisites: packs are plain Markdown skills. 2. Run `python install.py --dry-run` and show me the pack list and target (~/.claude/skills/jgs-nuclear-knowledge-packs/). If it looks right, run `python install.py`. 3. Verify: list the installed packs under the target and confirm the count matches SKILLS.md. 4. Other agents: `python install.py --list-agents`, then `python install.py --agent all` (claude, openclaw, copilot, codex, gemini). Shell: install.sh / install.ps1. 5. Tell me to restart the agent, then ask `/nuclear what NRC guidance covers software assurance for digital safety systems?` and check the answer cites pack + chapter. 6. Pack CONTENT is under each source's own licence (see NOTICE); keep attribution intact.
Every source clears a written eligibility rubric: Tier 1 public-domain or Tier 2/3 open-licensed with per-document capture where required. The Excluded list is a hard stop. Each pack carries its source's licence and attribution.
Each pack is an always-loaded SKILL.md index plus on-demand chapters, glossary, patterns, and a cheatsheet, so a large body of knowledge never fills the context window.
Plain Markdown. No MCP server, API key, or licence tier needed at runtime. Clone, install, invoke.
Stop breaking flow to dig through a long handbook. Ask a pack by slug mid-task and get the framework (with the chapter it came from) instead of a plausible-sounding guess. The agent answers from reconstructed reference notes, so you can cite the source, not a hallucination.
Put the same vetted bodies of knowledge in front of every engineer and every agent, junior or senior, on every project. Reviews, requirements, risk, and safety work trace back to open canon by default, not to whoever happened to remember it. Consistency you can audit, with no per-seat licence and no PDFs to chase.
Answers route to a specific chapter and name their source. The agent reads the text instead of recalling it, so you get fewer confident wrong answers in domains where that costs you.
Only genuinely open sources go in, each carrying its own licence. Hand a pack to a teammate, a subcontractor, or a client without a legal review.
Plain Markdown skills on the open Agent Skills format. No server, no API key, no subscription. They work the same offline as online.
# slash-invoke a pack by its slug
/<slug> # load the pack's core frameworks for reference
/<slug> <topic> # a topic: routes to the right chapter and answers
/<slug> ch06 # a chapter id: load that chapter directly
/<slug> what chapters do you have? # browse the index
# or just ask: each pack's description makes the agent load it on a match
"What does this body of knowledge say about verification?"The /nuclear orchestrator routes a free-text
Nuclear question to the right pack or packs, reads them, and answers with pack and chapter
citations. It carries no source content and answers only when explicitly invoked, so it pairs with every
pack in the catalogue.
/nuclear what does this pack say about risk acceptance? # narrow: routes one pack
/nuclear compare how two sources guide verification planning # broad: proposes a plan first
/nuclear draft the risk section of a plan # deliverable: pauses at each stageEach pack uses progressive disclosure: an always-loaded SKILL.md index decides which single chapter answers your question and loads only that, so a large source never fills the context window. Full detail in docs/skill-usage.md.
git clone https://github.com/jgsystemsconsulting/jgs-nuclear-knowledge-packs.git
cd jgs-nuclear-knowledge-packs
python install.py --dry-run # preview packs + target
python install.py # install for Claude Code
# other agents: native or format-transformed
python install.py --list-agents
python install.py --agent all # claude · openclaw · copilot · codex · geminiClaude Code, OpenClaw, and GitHub Copilot CLI read the SKILL.md format directly: the whole pack is copied, with full progressive disclosure.
Codex (prompt), Gemini (TOML command), and Cursor (project rule) receive the always-loaded index inlined into one file by the installer.
One stdlib-only install.py with --agent / --list-agents drives them all. Defaults to Claude Code.
That is 9 skills in all: 7 knowledge packs, 1 signpost, and the /nuclear orchestrator. Browse and filter the full catalogue on the
pack reference page, or read the raw index in
SKILLS.md.
MIT, by JG Systems Consulting Ltd. The installer, validator, docs, and repo structure.
Each pack carries its source's licence in packs/<slug>/LICENSE, independent of the repo MIT licence. A CC BY-NC-SA source yields a CC BY-NC-SA pack.
Every pack is reconstructed as synthesized reference notes, never a verbatim reproduction, with per-pack source attribution in NOTICE. No source-material download links are published; attribution travels as text plus the licence deed.