Claude Skills by nick-railsback

When the user wants to free disk space by deleting the skill-engine local clone cache (~/.cache/skill-engine/), with a dry-run preview before destructive action.

When the user wants the engine to discover the essence of registered sources and emit reference files. Suitable for first runs against a fresh contextualizer and for recurring rediscovery against a maturing one. The engine hands you a task and validates your output via the four reference invariants and verify.sh.

When the user is in an empty directory and wants to scaffold a new contextualizer from the bundled templates. Accepts one or more URLs or local paths to source roots; the engine auto-detects identity, kind, and topology — the user never types engine taxonomy.

When the user wants to add a single new reference file to an existing contextualizer without running the full discover pipeline.

When the user wants to refresh an existing contextualizer — re-check upstream state for registered sources, detect drift, and re-emit reference files where the upstream has changed. The engine hands you a task and validates your output via the four reference invariants and verify.sh.

When the user wants the engine to audit its own configuration and recently emitted artifacts for invariant violations and drift.

When the user wants a one-page report on the contextualizer's current health, recently emitted references, and pending review work.

When the user invokes any engine workflow from a contextualizer directory and the appropriate sub-workflow needs to be selected based on the directory's setup state.

Answers questions about nick-railsback/skill-engine — the Claude Code plugin that scaffolds and maintains 'contextualizer' skills. Covers its workflows (bootstrap, discover, refresh, review, apply, self-audit, status), the source-paths.json and reference-file artifact contract, the four reference invariants and verify.sh checks, the staging/proposed review model, monorepo handling, and versioning. References load on demand from references/.

Answers questions across the configured sources. References load on demand from references/. Use when working in any of this contextualizer's source domains.

Answers questions about the LangChain Python ecosystem — the langchain-ai/langchain monorepo (langchain-core, v1 agents, langchain-classic, in-tree partners, text-splitters) plus companion repos LangGraph, LangSmith SDK, Deep Agents, out-of-tree partner monorepos langchain-google and langchain-aws, and the docs.langchain.com source. Includes a thin orientation pointer for the LangChain.js port. References load on demand from references/.

Answers questions about the official Model Context Protocol (MCP) Python SDK (`mcp` package on PyPI, `modelcontextprotocol/python-sdk` on GitHub). Tracks the `main` branch (v2 pre-alpha — `MCPServer`/snake_case/constructor-`on_*` handlers). Use when working with MCP servers or clients in Python, debugging v1→v2 migrations, or answering questions about tools, resources, prompts, transports, OAuth, or experimental tasks. References load on demand from `references/`.

Use when a staged proposal has been reviewed and signed off — REVIEW.md Step 3 ticked reviewed or provisional — and is ready to promote into the live contextualizer.

Use when changing an engine-wide config value — currently `diff.tool`, the command `/skill-engine:review` prints for inspecting a proposed-vs-live diff.

Use when a staged proposal doesn't warrant promotion, or was ticked reject in REVIEW.md Step 3, to remove the proposed staging directory without touching the live contextualizer.

Use when DISCOVER or REFRESH has surfaced a proposal staged at <name>-context.proposed/, to inspect it and record sign-off before applying it.