Audits an AI/LLM codebase across nine surfaces using the SCAN method and returns a ship/conditional/hold verdict. Use when asked to review, audit, or security-check an AI agent, LLM app, RAG pipeline, or prompt/tool-calling system for AI-specific risks. Static analysis only.

Portable, advisory alternative to a hard read-size block. Canon installs zero Claude Code hooks — see Gotchas — so this skill cannot mechanically prevent a large direct `Read` the way a `PreToolUse` hook could. It generalizes `skills/sprint/reference/orient.md`'s citation-only subagent pattern i

Generates Marp slide decks from canon knowledge. Use when asked to create slides, build a presentation, or generate a deck on a canon topic (context management, evaluator pattern, or skill authoring). Renders to HTML and PPTX for browser viewing and SharePoint/PowerPoint distribution.

Audits always-on context load for bloat, redundancy, and quality. Use when context feels heavy or periodically to keep the always-on budget lean.

Finds code that nothing calls and reports it — it does not delete anything on its own. The scan edits no existing file; it writes a structured `dead-code-report.md` (confirm-first) and, only on per-candidate confirmation, removes code as a normal gated edit. Complements `mutation-test` (which chec

Plan a cascading refactor as a dependency graph you discover by *attempting* it, so the tree stays green at every step and nothing is half-migrated.

Before delivering any report or summary: define what a valid output looks like, generate, then verify the output against that spec.

Reviews LEARNINGS.md's UNPROMOTED rows as a fresh, no-implementation-history reader and proposes where each durable one belongs. Report-only, except that in a consumer project it writes the proposals you confirm to PROMOTED.md. Run it by hand, in a fresh session, when the UNPROMOTED queue needs triage.

Internal — not for direct/public invocation. Called by the maintainer when auditing canon's own workflow docs.