Render the EFFECTIVE Claude Code harness rulebook for this machine/project — what the current session may actually do — by merging every settings layer (global ~/.claude/settings.json, project .claude/settings.json, project settings.local.json) plus PreToolUse/PostToolUse hooks into one tiered view (free / consequential / guarded), flagging cross-shell asymmetries (a Bash rule with no PowerShell twin), wildcards that undo curated per-verb rules, and denies shadowed by allows. Use when you ask "what am I allowed to do", "what's my floor", "show the effective permissions", "render my rulebook", "why did that prompt", "audit my settings", "what's changed in my permissions", or invoke /floor. Read-only — it never edits settings.
Sign-on tool-health board — reads the workbench friction log and rolls it up per tool into a "how's each tool doing" view (recent friction, worst severity, one-line pain), using a LOCAL model (free, offline). Use when the operator asks "how are my tools doing", "tool health", "what's the friction", "how's everything doing", "anything need attention", "are my tools ok", or invokes /health, especially at sign-on. Distinct from /wip (in-flight work right now) and /status (session recap) — this is standing tool *health* from accumulated friction.
Offload a cheap, verifiable sub-task to the LOCAL model (Ollama, $0, offline) via the `local` CLI — narrow a big file list, extract structure from noisy tool output, classify log lines or command success/failure. Use PROACTIVELY mid-task when a sub-step is mechanical narrowing/extraction/shallow-classification AND you can verify the result at a glance; also when the operator says "offload that", "run that locally", "use the local model", or invokes /offload. NEVER for deep judgment (code review, risk calls, dense-diff reasoning) — that stays with you.
Build and mine a corpus of AI-draft → operator-final pairs so the agent's writing converges on the operator's real voice. Two verbs — capture (freeze the AI version of a draft BEFORE the operator edits it) and learn (diff the operator's final against the frozen AI version, classify every edit as voice vs content, and fold recurring voice edits into voice-profile.md and tellscan.mjs as dated, evidence-backed rules). Use when the operator says "voice learn", "capture this draft", "here's my final version, learn from it", "diff my edits", "store this pair", "learn my voice", or invokes /voice-learn. Corpus lives under a writing dir you configure (e.g. ~/writing/corpus/); the voice profile stays the single contract.
Fork a thread that came up mid-task into its own briefed Claude Code session — without derailing the agent that's currently working. The source agent pauses just long enough to emit a thread-scoped context packet, spins a chip seeded with it, and resumes its main task untouched. The chip becomes a fresh session you can question about that one thread. Use when the user says "transfer context", "pull on that thread", "fork this thread off", "spin that tangent into its own session", "hand this thread to another agent", "don't derail — chip that thread with context", or invokes `/transfer-context`. Distinct from /chip (fires off a TASK, and explicitly avoids conversation-dependent context) and /continue (hands off the WHOLE session). This is thread-scoped, conversation-dependent, and non-disruptive to the source.
Skill librarian - catalog every skill available in this session and recommend the best fit for a task. Discovers all skills on disk (personal ~/.claude/skills, project .claude/skills, and skills stranded in .claude/worktrees) by reading their frontmatter, so the catalog is always current. Run alone (/skills) for a grouped tabular overview of everything you have; run with a prompt (/skills open a pull request for my changes) to rank the catalog against that task and recommend the top matches with the exact command to type. Use when the operator asks "what skills do I have", "which skills are available", "what can I run", "is there a skill for X", "which skill should I use for Y", "recommend a skill", "I forget what skills I built", "what's available to me", or invokes /skills.
Consolidate the AI PR reviewers (Codex, Claude, Cursor Bugbot, GitHub Copilot) on a pull request into ONE verdict — a deduped, severity-ranked finding list plus a block/go merge gate — and post it as a single sticky comment layered on top of (never replacing) the bots' own comments. The "judge" over the finders: it consolidates, it never produces findings of its own. Use when you (or work-driver) want one pane + one merge decision instead of reading four separate bot comment streams. Triggers: "consolidate the reviews on PR <n>", "what's the verdict on PR <n>", "judge the bot findings", "/review-coordinator <n>". NOT a reviewer itself — pair it AFTER the bots have reviewed.
Turn a PR's AI-reviewer comment pile (codex / cursor / claude / copilot) into a clean, line-grouped digest using a LOCAL model — each bot's own headline + stated severity + verdict, extracted (not re-judged), grouped by file:line. Free, offline, no cloud tokens. Use when someone says "triage the comments on PR N", "what did the bots flag on PR N", "digest the review comments", "summarize the PR feedback", or invokes /review-digest N. A pre-pass that feeds a review coordinator — advisory only, it never gates a merge.