القائمة
Guide for creating Claude Code skills from scratch. Use when building new skills, understanding skill file structure, writing effective descriptions, or improving existing skills.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Full-stack Hermes Agent health audit: observability config, compression tuning, prompt/skill consistency, skill invocation diagnosis, system overload detection, session pattern analysis, tool-use/delegation tuning, plus a full annotated config (model/queue/IO/security), per-channel Feishu output discipline, a prompt-engineering method, and a reverse-QA regression loop. Use when skills stop triggering, model ignores loaded skill steps, context grows unexpectedly, sessions hit compression too often, long tasks get interrupted, the task queue misbehaves, or after editing any prompt/skill/config file.
Securely serve private S3 content through CloudFront using Origin Access Control (OAC). Use when you need to make S3 files publicly accessible over a CDN without making the bucket itself public — e.g. hosting static sites, media, downloads, or build artifacts. Covers the modern OAC approach (AWS-recommended), migrating legacy OAI distributions to OAC, locking down the bucket Public Access Block, enforcing HTTPS, and verifying the result. Never make an S3 bucket public directly.
用 AWS Bedrock AgentCore 云端浏览器做联网深度调研、搜索、抓网页正文。Use when 要做带引用的多源调研 / 搜资料 / 抓动态页 / 反爬站 / SPA / 需 JS 渲染的页面。本机已部署同名 MCP server，优先用它而非裸 WebFetch。NOT for 抓本机已登录态的页面(用 chrome-cdp 9222)，NOT for 纯本地文件搜索(用 grep/search_files)。
用本机已登录的 Perplexity / Gemini 做 AI 深度搜索,直接拿带引用的综合报告。当需要快速拿一个话题的多源综述、对比、"现在最好的 X 是什么"、带引用的事实核查,且想省去自己拼 web_search+抓取时使用。走 CDP 驱动本机 Chrome(9222,已登录 Pro)。NOT for 抓指定单个 URL 的正文(用 fetch_page),NOT for 需要自己控制每个来源的严谨调研(用 deep-search-protocol + agentcore-deepsearch)。
Desktop notifications for Claude Code via Notification/Stop hooks. Pops a native macOS notification showing which project, who's waiting, and what for — with click-to-focus back to your terminal. Use when setting up sound alerts, desktop notifications, or 'tell me when Claude needs me / is done'.
主动管理 context 质量与信噪比，确保决策基于可信、新鲜、相关的信息
| name | skill-creator |
| description | Guide for creating Claude Code skills from scratch. Use when building new skills, understanding skill file structure, writing effective descriptions, or improving existing skills. |
A practical guide to creating high-quality Claude Code skills. Covers file structure, description writing, common pitfalls, and testing.
skills/
└── my-skill/
├── SKILL.md # Required — the skill definition
├── references/ # Optional — supporting docs, examples
│ ├── example.md
│ └── patterns.md
├── agents/ # Optional — sub-agent definitions
│ └── specialist.md
└── scripts/ # Optional — automation scripts
└── scan.py
---
name: my-skill
description: "One-line description. Use when [trigger conditions]."
metadata: # Optional
openclaw:
emoji: "⚡"
requires:
bins: ["tool1", "tool2"]
---
# Skill: my-skill
Brief intro — what this skill does and why it exists.
## When to Use
Trigger conditions and decision criteria.
## Quick Start
Minimal steps to get value immediately.
## Core Content
The actual methodology, workflow, or reference material.
## Anti-Patterns / Gotchas
Common mistakes and how to avoid them.
The description field is the most important line in your skill. It determines when the skill gets activated.
| Bad | Good |
|---|---|
| "A tool for helping with code" | "Code quality review with security focus. Use after writing code, before merging." |
| "Handles AWS things" | "AWS infrastructure queries via CLI. Use for auditing, monitoring, and resource inventory." |
| "Testing skill" | "E2E browser testing via Chrome DevTools. Use when verifying UI flows, form submissions, or visual regressions." |
A skill should do one thing well. If your SKILL.md covers both "database optimization" and "API design," split it into two skills.
Users need to know WHAT to do, in WHAT order. Structure your skill as a numbered workflow, not a reference manual.
When the skill involves choosing between approaches, provide a decision matrix:
| Scenario | Approach | Why |
|----------|----------|-----|
Define what good output looks like. Include templates with placeholders that the agent can fill in.
If there are non-negotiable constraints, state them explicitly in a "Hard Rules" or "Constraints" section. Agents follow explicit rules better than implicit conventions.
A "Don't do this" section prevents the most common mistakes. Format as a table: Don't | Do Instead.
Large lookup tables, CLI command references, or pattern libraries belong in references/, not in SKILL.md. Keep the main file focused on workflow.
Before shipping, test your skill with 5+ realistic user prompts. Check:
Skills are plain Markdown — use git for versioning. Commit messages should explain WHY the skill changed, not just what changed.
| Pitfall | Impact | Prevention |
|---|---|---|
| Description too vague | Skill never activates | Include specific trigger phrases |
| SKILL.md too long (>500 lines) | Agent loses focus mid-skill | Extract to references/ |
| No output format defined | Inconsistent results | Add a template with placeholders |
| Hardcoded paths or secrets | Skill breaks on other machines | Use environment variables or relative paths |
Missing When to Use section | Users don't know when to apply it | Always include trigger criteria |
| Overly rigid workflow | Doesn't adapt to variations | Mark optional steps clearly |
| No anti-patterns section | Users repeat common mistakes | Add "Don't / Do Instead" table |
Test your skill systematically:
name is lowercase-kebab-casedescription includes "Use when" triggerreferences/ if SKILL.md > 300 lines