본문으로 건너뛰기
cccskills
무료GitHub에 공개

my-agent

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.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

연동 도구OpenAIGemini
설치 방법 보기

포함된 파일(4)

  • SKILL.md3.8 KB
  • references/claude.md2.9 KB
  • references/codex.md1.6 KB
  • references/gemini.md2.9 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

CLI Agent Runner

Delegate tasks to an external CLI agent from within Claude Code. Supports three providers plus Claude's explicit fable model alias:

  • codex — OpenAI Codex CLI (codex exec ...)
  • gemini — Google Antigravity CLI (agy -p ...)
  • claude — Claude Code CLI (claude -p ...)
  • fable — Claude Code CLI with explicit model alias (claude --model fable -p ...)

Provider Selection

Pick the provider in this order:

  1. User says "fable" / "Fable" / "fableに相談" / "Fableに聞いて" / asks to delegate or consult Claude Fable → claude with --model fable
  2. User says "gemini" / "ジェミニ" / asks to delegate to Gemini → gemini
  3. User says "claude" / "クロード" / asks to delegate to Claude → claude
  4. User says "codex" / "コーデックス" → codex
  5. No explicit provider mentioned (default) → codex

The first argument is treated as a provider only if it is exactly codex, gemini, or claude. If the first argument is exactly fable, use the claude provider with --model fable and treat the remaining arguments as the task prompt. Otherwise the entire argument string is the task prompt and codex is used. Do NOT ask the user which provider to use — pick codex and proceed.

The remaining arguments form the task prompt. For consultation phrasing such as "fableに相談して X" or "Fableに聞いて X", remove the consultation phrase and use X as the task prompt. Convert the user's request into a clear, self-contained prompt and include relevant file paths/context. If the user provides a literal prompt, pass it through directly.

Common Notes (all providers)

  • Always use the shell command tool with a 10-minute timeout; delegated tasks can take time.
  • Do NOT hardcode model names as defaults — let each CLI's own default handle it so it stays current. Pass --model fable only when the user explicitly requests Fable.
  • Output streams to stdout; capture or display results directly.
  • For long-running tasks, warn the user that the delegate may take time.
  • If the delegate fails, report the error and suggest the user run it interactively (codex / gemini / claude).

Quick Reference

Codex

codex exec "<PROMPT>"

For image input, model override, resume, configuration details, and more examples, read references/codex.md.

Gemini

Before the first agy -p of the session, read references/gemini.md. It contains the mandatory preflight, sandbox, and headless-mode requirements.

agy --sandbox --disable-slash-commands -p "<PROMPT>"

For execution modes, structured output, model override, troubleshooting, and more examples, read references/gemini.md.

Claude

Before the first claude -p of the session, read references/claude.md. It contains the non-interactive invocation rules and permission-mode guidance.

claude -p "<PROMPT>"

# With explicit Fable model alias
claude --model fable -p "<PROMPT>"

For permission modes, model override, structured output, troubleshooting, and more examples, read references/claude.md.

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

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.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Ynakatsuka/dotfiles42026년 10월 18일 업데이트

my-ccv

무료

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.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Ynakatsuka/dotfiles42026년 10월 18일 업데이트

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.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Ynakatsuka/dotfiles42026년 10월 18일 업데이트

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.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Ynakatsuka/dotfiles42026년 10월 18일 업데이트

my-epic

무료

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.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Ynakatsuka/dotfiles42026년 10월 18일 업데이트

Hand off a task with only the context needed to continue in a fresh session. Use for "handoff", "引き継ぎ", "別セッションへ渡す", or "別マシンで続ける". The orca subcommand launches an independent Orca workspace or terminal tab; without it, produce a paste-ready prompt backed by a verified remote branch or PR. Do not launch sessions for capability questions or design requests.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Ynakatsuka/dotfiles42026년 10월 18일 업데이트

Ynakatsuka의 스킬 모두 보기

이 스킬의 문제 신고