Iron rule: Do not recommend a skill just because a keyword appears in the search results. First confirm the problem the user actually wants to solve, then verify the source, quality, and installation cost.
Workflow
- Step 1: Clarify what capability the user wants to extend ⚠️ REQUIRED
- 1.1 Extract the domain, task, inputs and outputs, and constraints.
- 1.2 Determine whether the user needs "a skill already in this repository" or "an external ecosystem skill".
- Step 2: Check the current repository first
- 2.1 Review the existing skills in AGENTS.md and the skills/ directory.
- 2.2 If an existing skill matches well enough, prioritize recommending the local skill over an external installation.
- Step 3: Search the public skill ecosystem (conditional)
- 3.1 First check the skills.sh leaderboard to confirm whether a high-install-count skill already exists.
- 3.2 When a more detailed search is needed, run npx skills find .
- Step 4: Verify candidate quality ⚠️ REQUIRED
- 4.1 Check install count, author reputation, repository activity, and GitHub stars.
- 4.2 Filter out skills with unknown sources, extremely low install counts, or vague purposes.
- Step 5: Give recommendations
- 5.1 Explain the skill name, purpose, source, install count, and why it matches.
- 5.2 Provide the installation command, skill page, and possible alternatives.
- Step 6: Handle the no-results scenario
- 6.1 Clearly state that no suitable skill was found.
- 6.2 Provide a direct assistance plan, or suggest using skill-creator to create a new skill.
Common commands
npx skills find <query>
npx skills add <owner/repo@skill>
npx skills add <owner/repo@skill> -g -y
npx skills check
npx skills update
Recommendation order
- First check whether a reusable skill already exists in this repository.
- Then look at skills with high install counts and trustworthy sources on skills.sh.
- Only recommend "create a new skill yourself" as the last option.
Quality threshold
- For install count, prefer skills with 1K+, and when below 100, clearly flag the risk.
- Prefer trustworthy sources, such as anthropics, microsoft, vercel-labs, etc.
- If the search results come from a low-activity repository or have vague descriptions, they cannot be recommended directly.
- Before installation, explain what capability the skill adds and whether it overlaps with existing skills.
Recommendation output template
我找到了一个可能匹配的技能：<skill-name>
- 用途：<这个技能能解决什么问题>
- 来源：<owner/repo>
- 可信度：<安装量 / 仓库情况 / 是否官方来源>
- 安装命令：npx skills add <owner/repo@skill>
- 了解更多：<skills.sh 或仓库链接>
如果你希望，我可以继续帮你安装，或者先比较几个候选项。
Anti-patterns
- Guessing the purpose from the skill name alone without reading the description and source.
- Directing the user to install an external skill without first checking the current repository.
- Presenting low-quality or unknown-source skills as "recommended options".
- Not explaining the benefits, risks, and alternatives after installation.
Pre-delivery checklist
- The user's task has been clarified, rather than just latching onto surface keywords.
- The repository skills and AGENTS.md have been checked first.
- Each candidate includes its source, quality, and installation command.
- When there are no results, a direct assistance or skill-creation alternative path has been provided.