Claude Code Skills Best Practices

功能对比

Keyword Evidence

claude code skills best practices is a P1 supporting keyword for the main Claude Code Skills page. Latest backend data provided for this cluster shows US Volume 110, CPC $12.41, Global Volume 220, keyword ideas 8, and KD unavailable.

Best Practices Checklist

Treat each skill as a small operating procedure for an agent. It should help Claude choose the right workflow, inspect the right context, produce a predictable output, and stop before risky actions.

• Name the workflow in plain language, such as code-review-skill or docs-refresh-skill.
• Write the description as a trigger, not as a marketing summary.
• Keep required inputs, allowed context, and forbidden actions explicit.
• Add a short validation path so the result can be checked.
• Move bulky examples into references/ or templates/ instead of inflating SKILL.md.

Skill File Structure

A skill can be just one SKILL.md file. Add supporting files only when the workflow needs repeatable references, scripts, or output templates.

.claude/
  skills/
    code-review-skill/
      SKILL.md
      references/
        review-checklist.md
      templates/
        review-output.md

# Personal reuse:
~/.claude/skills/code-review-skill/SKILL.md

When To Use Skills, Hooks, Or Commands

The extension surface should match timing. Skills are reusable workflow knowledge, hooks are deterministic automation, and commands are explicit user-triggered shortcuts.

• Use skills for repeatable reasoning workflows with examples, references, or templates.
• Use hooks for automatic lifecycle checks such as linting after edits or blocking risky tool use.
• Use slash/custom commands for short workflows a developer wants to trigger manually.
• Use MCP when the agent needs external data or tool access beyond repo files.

Copyable Skill Template

Use this starter template for a team skill. Replace placeholders, keep tool access narrow, and remove any example that contains private data.

---
name: <skill-folder-name>
description: Use this skill when <specific trigger and task boundary>.
---

# <Skill Name>

## When To Use
Use this skill for <repeatable workflow>. Do not use it for <out-of-scope work>.

## Inputs
- Repository or files to inspect:
- User goal:
- Constraints, deadlines, or style rules:

## Workflow
1. Confirm the task matches this skill.
2. Read only the files or references needed for the workflow.
3. Identify risks, assumptions, and missing context.
4. Produce the requested artifact or edits.
5. Run or recommend the validation command.

## Output Format
Return:
- Summary
- Findings or changes
- Validation result
- Follow-up risks

## Safety
- Do not print secrets, tokens, private customer data, or billing details.
- Do not run destructive commands unless the user explicitly approves.
- Keep network, deployment, and database access behind explicit approval.

Security Notes

Skills can encode habits that run repeatedly, so security review matters even when the skill is only Markdown.

• Do not put API keys, credentials, account IDs, or private exports in SKILL.md.
• Avoid broad Bash patterns in review-only skills.
• State when a human must approve write actions, deploys, purchases, or data deletion.
• Review third-party skill examples before copying them into a repo.
• Keep customer data and internal URLs out of public template examples.

Review Checklist

Before sharing a skill with a team, review the trigger, scope, tools, data exposure, and verification path.

[ ] The skill has a specific trigger description.
[ ] The workflow is narrow enough to test on one task.
[ ] No secrets, tokens, private URLs, or customer data are included.
[ ] Commands are scoped and non-destructive by default.
[ ] The output format is predictable.
[ ] The validation command or manual check is listed.
[ ] The skill links back to CLAUDE.md or the team docs.

Sources And Next Pages

Use official Claude Code docs for skills behavior, then use the main skills directory and agent skills directory to choose where the workflow belongs.

常见问题