AgentSkillsCN

claude-code-coagent

将 claude_code MCP 用作第三方协作者,收集替代意见、打破僵局或委派特定任务(评审、想法、草稿编辑),同时你仍保持决策权。当寻求另一种视角或备用执行人时触发:“其他意见”、“另一个人”、“不同观点”、“没有突破口?”、“另一个 LLM 的见解”、“请 Claude 帮忙”。

SKILL.md
--- frontmatter
name: claude-code-coagent
description: >
  Use claude_code MCP as a third-party co-agent to gather alternative opinions, unblock stalemates,
  or delegate scoped tasks (reviews, ideas, draft edits) while you remain the decision-maker.
  Trigger when seeking another perspective or backup executor: "他の意見", "別の人", "違う観点",
  "打開策ない?", "別LLMの見解", "claudeに頼んで".

Claude Code Co-Agent

What this skill is for

  • Pull in claude_code as an external reviewer/ideator/executor while you (Codex) keep ownership.
  • Typical asks: alternative solutions, risk review, unblock ideas, scoped edits/tests in bounded paths.
  • Use when the main pair (user + Codex) is stuck or wants a second opinion or fast trial edit.

Roles

  • You (Codex): own context, guardrails, final decisions, and integration of any changes.
  • claude_code: provides opinions, options, or scoped file edits; must work within explicit bounds.

Workflow

  1. Frame the ask
    • Goal and expected output format (e.g., 3 options ranked, risk bullets, patch for file X only).
    • Scope: allowed files/dirs, banned actions, time/complexity constraints.
    • Context: minimal code snippets/diffs needed to answer; avoid oversharing.
  2. Invoke claude_code with a structured prompt (see templates below).
    • タイムアウトしたら1回リトライ。成功しない場合はユーザーへ報告。
  3. Receive and digest
    • Summarize response in 2–3 bullets。
    • MUST: 返答内容を自分の結論・回答に反映する。統合しない場合は理由を明記。
    • Decide: adopt / partial / discard.
    • If edits returned, review diff → apply → run relevant checks/tests if riskful.
  4. Follow-up if needed
    • Ask to narrow to one option, or request a patch limited to specific paths.
  5. Record briefly
    • What was asked, key takeaways, what was applied (optional but recommended for continuity).

Prompt templates

  • Opinion/ideas: "目的: <goal>. 制約: <constraints>. 求める出力: 選択肢3件を重要度順で。禁止: <bans>."
  • Review: "次のdiffを重大リスクがある点のみ箇条書きで。無害な指摘は不要。<diff/snippet>"
  • Scoped patch: "ファイル限定: <paths>. 仕様: <spec>. 期待: 完全なパッチのみ。禁止: 依存追加/構造変更。"
  • Unblocker: "今の案が行き詰まり。打開策3件、各1行で。前提: <brief>."

Safety/guardrails

  • Always state allowed paths and banned actions before asking for edits.
  • Prefer getting reasoning or options first; request patches only after scoping.
  • Verify diffs yourself; run focused tests when behavior can change.
  • タイムアウト時は1回リトライし、成功した返答は必ず要約して結論に反映。統合しない場合は理由を明示。リトライ失敗時はユーザーへ報告。

References

  • Prompt examples: references/prompts.md
  • Call checklist: references/checklist.md