Thanks for helping build this out.

• Keep SKILL.md lean. It should stay well under ~500 lines and route to references/.
• Depth goes in references/. One file per domain; add a new file rather than bloating an existing one.
• Ground claims. Cite the current code edition or standard; note when a value must be locally verified.
• Universal vs local. Separate transferable reasoning from jurisdiction-specific values.

• Country / jurisdiction dossiers (code family, AHJ, load basis, licensure).
• Additional worked case studies under examples/.
• Corrections to code-edition or standard references as cycles update.

1. Fork and branch.
2. Make the change; if it touches triggering, update the description in SKILL.md frontmatter.
3. If you added a capability, update its row in the SKILL.md reference table. That table is the runtime routing map — it is what Claude reads to decide which reference to open, and it is also the source the MCP server's list_references serves. The retrieval eval scores reference bodies, so a stale table passes CI silently. This one is on the author.
4. Rerun the build so the packages and the portable bundle stay in sync with the source:

   python scripts/build.py

   Everything in dist/ and plugin/skills/master-builder/ is generated — never edit those by hand. If you add or rename a reference, the build picks it up automatically; just commit the result.
5. Check your work (CI runs exactly these, on Python 3.9 and 3.12):

   python scripts/validate.py            # structure, links, and the plugin manifests
   python scripts/eval_retrieval.py      # every question still routes to the right reference
   python scripts/eval_behavior.py --check
   python scripts/mcp_server.py --selftest

   If you add substantial new material, add a retrieval case for it in evals/retrieval.jsonl — a topic the eval doesn't cover can silently rot.
6. Open a PR describing what changed and why.

Every script in scripts/ is stdlib-only — no dependencies, no lockfile, nothing to install. Keep it that way.