Guidance for editing Claude Code agents. Use when Claude touches agent markdown files so tool scopes, MCP wiring, memory scope, isolation, and agent hooks are configured safely.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Guidance for editing Claude Code agents. Use when Claude touches agent markdown files so tool scopes, MCP wiring, memory scope, isolation, and agent hooks are configured safely.
Agent Authoring
When editing agents:
Prefer explicit tools lists for read-only agents. Use wildcard plus disallowedTools only when you truly need the broad tool pool.
Use mcpServers to add docs/search tooling without hardcoding MCP tool names into the agent prompt.
Use initialPrompt for short steering that should apply on every spawn; keep the main prompt focused on durable review or execution behavior.
memory: project is for reusable repo knowledge. memory: local is for machine-specific workflows that should not leak across environments.
background: true is for long-running monitors or validators that do not need to block the main thread.
Agent hooks are session-scoped to the agent lifetime. Stop hooks become SubagentStop, so keep them outcome-oriented and lightweight.