AgentSkillsCN

claude-code

通过 `claude` 命令使用 Anthropic Claude Code CLI 来完成编码任务、代码重构、调试、测试以及仓库范围内的编辑。当用户请求使用 Claude Code、说出“用 claude/claude code”，或希望将某项任务委托给本地的 Claude CLI 时，即可触发该功能。

SKILL.md
--- frontmatter
name: claude-code
description: Use Anthropic Claude Code CLI via the `claude` command for coding tasks, refactors, debugging, tests, and repo-wide edits. Trigger when the user asks to use Claude Code, says "用 claude/claude code", or wants a task delegated to the local Claude CLI.

Claude Code CLI Skill

Use the local Claude Code CLI through exec.

Core rule

  • •Always invoke Claude Code with the claude command (not other wrappers).

Fast workflow

  1. •Confirm CLI availability:
    • •which claude
  2. •If missing, report clearly and stop.
  3. •Run Claude task with one-shot prompt in the project directory.
  4. •Return concise result + changed files + next step.

Command patterns

One-shot task

bash
claude -p "<task prompt>"

Safer deterministic run (recommended)

bash
claude -p "<task prompt>" --output-format text

Include repo context in prompt

State exact constraints in prompt (tests, style, no breaking changes, files to touch).

Example prompt text:

  • •"Fix failing tests in api/, keep behavior unchanged, then summarize modified files."
  • •"Refactor this function for readability without changing output; preserve type hints."

Execution guidelines

  • •Run in the intended repo root via exec.workdir.
  • •For long tasks, run in background and stream logs via process.
  • •After Claude finishes, validate quickly (run tests/lint if requested).
  • •Do not claim success without command output.

Failure handling

  • •If claude not found: suggest installation/login and stop.
  • •If auth/session errors occur: ask user to re-auth in terminal, then retry.
  • •If task is ambiguous: ask one focused clarification question.