Use when the user wants to run a screen-recordable end-to-end walkthrough of every public-facing sorcery skill — typical phrasings "run the sorcery demo", "let's record the demo", "walk through every skill", "demo all the skills".
Use when the user wants Claude Code to launch with privacy-friendly defaults (hidden account email/org, --rc, --effort max, --model 'claude-opus-4-8[1m]') for a specific repo. Creates an executable `claude.sh` at the repo root so the user runs `./claude.sh` locally — no shell aliases, no symlinks, no $PATH changes.
Use when the user wants a quick-start setup for routing a coding agent through many LLM providers — phrasings like "set up LiteLLM", "I want to try DeepSeek/Kimi/Qwen/MiniMax", "let me switch between OpenAI and Claude and Hugging Face from one place", "set me up with Swival", "route my coding agent through multiple models", "experiment with cheap vs frontier LLMs", "bootstrap a multi-provider LLM setup". Scaffolds an `llm-routing/` directory plus repo-root entrypoints (`llm.sh`, `swival.toml`, `.env.example`), with LiteLLM as the OpenAI-compatible router and Swival as the coding agent. Toolchain is pinned via a Nix flake; `setup.sh` auto-installs Nix via the Determinate installer when missing.
Use when the user wants to install an autonomous improvement loop into the current repo — a harness that launches Claude Code repeatedly (one iteration at a time) with rotating personas to find and fix issues unattended. The harness ships with a four-subprocess watchdog, a soft-deadline wrap-up hook, and changelogs that persist learning across iterations.
Use when the user wants to run Claude Code inside an isolated macOS VM (sandboxed from the host) — phrasings like "set up Claude in a VM", "isolate Claude", "give Claude a macOS sandbox", "run Claude without touching my host". Creates a subdir (default `claude-vm/`) with Tart-based setup scripts that clone a macOS image, boot it with VNC/Screen Sharing, and install Claude Code plus a small set of nice-to-have apps. Apple Silicon only.
Use when the user wants to install a pre-commit dependency-staleness check ("enforce periodic upgrades", "install the staleness hook", "nag me to bump deps", "block commits when bun.lock is old"), when the existing hook reports `[periodic-upgrades] Stale dependencies detected`, or when the user wants to do an open-ended dependency-refresh sweep ("make this repo current", "bump everything"). Installs a self-contained `.githooks/check-update-staleness.sh` and walks through the backup → upgrade → test → commit cycle.
Use when the user wants to set up a new repo with the sorcery day-one essentials in one shot — phrasings like "scaffold this repo", "bootstrap this repo", "initialize this repo with sorcery defaults", "set up the basics in this repo", "install everything". Bundles the universal-baseline installers behind one entry point and hands off to `following-best-practices` for any remaining gaps.
Use when the user wants to capture real-world web pages as test fixtures and store them with provenance — phrasings like "capture this page as a fixture", "set up test fixtures from real URLs", "snapshot this page for tests", "build a fixture from the live site", "I need a fixture for the parser tests". Bundles `capture.sh` (curl or Playwright-rendered DOM) and `strip.ts` (mechanical noise-stripping via Bun's HTMLRewriter) under the plugin's `fixtures/` directory. Codifies a two-pass simplification model — the script does the mechanical strip, then the LLM does a test-aware semantic trim — so every fixture is small, fast to load, and answers "where did this come from?" / "is it stale?" at a glance.