Delegate tasks to an external CLI agent (OpenAI Codex, Google Gemini, or Claude Code) for second opinions or parallel execution. Use when the user explicitly mentions "codex", "gemini", or "claude" or asks to delegate work to one of them (e.g., "codexで実行", "codexに聞いて", "geminiで実行", "Geminiに聞いて", "claudeで実行", "Claudeに聞いて", "run with codex", "ask gemini", "ask claude"). Also use when the user mentions Claude's "fable" model alias for delegation or consultation (e.g., "fableで実行", "fableに相談", "Fableに聞いて", "ask fable"). Do NOT use for general coding tasks that don't mention codex, gemini, claude, or fable.

Delegate reading large or multiple files to the `bulk-read` command, which sends the files to a cheaper reader model and returns only its answer. Use when a question needs the content of a file over about 350 lines or of several files at once, when a PreToolUse guard denied an unbounded read, or when the user says "bulk read", "一括読み込み", or "まとめて読んで". Do NOT use for files you are about to edit, for subtle correctness reasoning such as concurrency bugs, or for short files that fit a bounded read.

Manage Claude Code Viewer (CCV): preview in-progress CCV changes on ccv-preview, register HTML artifacts, download files, or schedule jobs via CCV local API. Use when user asks for CCV preview, ccv-preview, browser visual QA, UI/UX validation, CCV artifacts, CCV file download, or CCV scheduler jobs. Treat preview/プレビュー as the CCV app preview service first; use artifacts only when the user explicitly asks for artifact/dashboard/report HTML. Do NOT use for generic web app previews outside CCV, production deployment, Cloudflare config, or Anthropic remote triggers.

Delegate writing predictable code to the `code-write` command, which has a cheaper writer model write the target file directly so the generated code never enters your context. Use for tests, scaffolding, fixtures, boilerplate, or a new file that copies an existing pattern, and when the user says "code write", "生成を委譲", or "雛形を書かせて". Do NOT use for design decisions, changes to public contracts or shared modules, edits inside an existing file that needs judgment, or logic with subtle correctness requirements.

Disciplined diagnosis workflow for hard, intermittent, or multi-component bugs and performance regressions: inspect logs and existing evidence, build a feedback loop, test falsifiable hypotheses, instrument gaps, and verify the outcome. Use when the root cause is unclear, reproduction is unreliable, several causes or components are plausible, or performance has regressed. Do NOT use for obvious single-cause errors, straightforward configuration or syntax fixes, planned feature work, or broad architecture review. Diagnosis-only requests stop after reporting the root cause and evidence; apply a fix only when the user requested one.

日本語の概要は準備中です。原文の説明を表示しています。

Manual-only orchestrator for medium-to-large software development. Turns a broad goal into PR-sized and operational nodes with verification gates, then coordinates approved execution across follow-up prompts. Activate only when the user explicitly invokes my-epic; once activated, keep orchestrating that epic for the rest of the session. Do NOT use for an uninvoked planning or implementation request, a single small bug fix, one obvious PR, pure code review, ordinary SDD work that fits in one spec/PR, or implementing an epic during the same first call that created it.

日本語の概要は準備中です。原文の説明を表示しています。