AgentSkillsCN

claude

当用户提出“使用 Claude Code”“Claude Code 最佳实践”“无头运行 Claude”“claude -p”“计划模式”“配置 Claude Code”或“优化 Claude Code 工作流”等需求时，应使用此技能。

SKILL.md
--- frontmatter
name: claude
description: This skill should be used when the user asks to "use Claude Code", "Claude Code best practices", "run Claude headless", "claude -p", "plan mode", "configure Claude Code", or "optimize Claude Code workflows".

Claude Code Best Practices

Use Claude Code as an agentic environment. Emphasize verification, context discipline, and structured workflows.

When to Use

  • •Onboarding or workflow questions about Claude Code
  • •Designing prompts for multi-step changes
  • •Headless automation (claude -p) and CI usage
  • •Scaling via subagents, hooks, MCP, skills, or parallel sessions

Core Principles

  • •Provide verification criteria (tests, screenshots, expected output).
  • •Explore → Plan → Implement → Commit; skip planning for tiny edits.
  • •Scope prompts tightly (file paths, constraints, patterns, edge cases).
  • •Manage context aggressively; avoid kitchen-sink sessions.
  • •Fix root causes; avoid suppressing errors.

Workflow Pattern

  1. •Explore in Plan Mode: read files, answer questions, no edits.
  2. •Plan: list file changes, risk areas, and checks.
  3. •Implement: make edits and verify against the plan.
  4. •Commit/PR: only when explicitly requested.

Prompting Patterns

  • •Reference files with @path/to/file instead of describing locations.
  • •Include examples and expected outcomes; supply failing cases.
  • •Describe symptoms and "fixed" criteria.
  • •Ask for an interview using AskUserQuestion for large features.
  • •Paste screenshots/images for UI work and specify what to match.
  • •Provide URLs for docs/APIs; allowlist domains via /permissions when needed.
  • •Provide data via pipes: cat error.log | claude.

Headless Mode

  • •One-off query: claude -p "Explain this project"
  • •JSON output: claude -p "List API endpoints" --output-format json
  • •Streaming: claude -p "Analyze logs" --output-format stream-json
  • •Stream event types: init, message, tool_use, tool_result, error, result.
  • •Batch/fan-out: loop through files and use --allowedTools to scope actions.
  • •Use --verbose for debugging.
  • •Use --dangerously-skip-permissions only inside a sandbox without internet.

Environment Setup

  • •Keep CLAUDE.md concise; include only non-obvious rules and commands.
  • •Use @ imports in CLAUDE.md for related docs.
  • •Configure permissions with /permissions or isolation with /sandbox.
  • •Prefer CLI tools (gh, aws, gcloud, sentry-cli) over API calls.
  • •Connect MCP servers with claude mcp add.
  • •Use /plugin to discover and install plugins.
  • •Use hooks for must-run actions; configure with /hooks or .claude/settings.json.
  • •In this kit, hooks/stop.sh provides deterministic autonomous-loop enforcement based on .claude/autonomous-loop.json (git clean + task list complete + quality gates + scoped plan checks before exit).
  • •Create skills in .claude/skills/ and subagents in .claude/agents/.

Context Management

  • •Use /clear between unrelated tasks.
  • •Use /compact <instructions> to guide compaction.
  • •Use Esc to stop runaway actions while preserving context.
  • •Use /rewind or double-ESC to restore checkpoints.
  • •Resume with claude --continue or claude --resume; rename with /rename.

Parallelism

  • •Run parallel sessions for writer/reviewer workflows.
  • •Use subagents for investigations or second-pass reviews.

Common Pitfalls

  • •Kitchen-sink sessions: reset with /clear.
  • •Repeated corrections: restart with a cleaner prompt.
  • •Overlong CLAUDE.md: prune until only essential rules remain.