How to Use Skills

This document explains how to use CloudBase AI Skills in projects. Skills are "skill packages" containing instructions and resources that inject CloudBase professional knowledge and steps into AI; when used with MCP, MCP provides connections and permissions, while Skills provide rules and intuition, resulting in better outcomes.

Method 1: Install Skills (Recommended)

Install CloudBase Skills to your local project via the Skills Open Specification, suitable for AI tools that support Skills (such as Claude Code, Cursor, VS Code, GitHub Copilot, OpenCode, etc.).

Install Skills

Execute in the project root directory:#!

npx skills add tencentcloudbase/cloudbase-skills

Select your AI tool according to the prompts. After installation, Skills will appear in the current project (e.g., .cursor/skills, .claude/skills, etc., depending on the tool).

If you want to trigger installation directly via natural language, you can also tell AI "install CloudBase Skills".

If you only want to install a single Skill, you can execute:#!

npx skills add https://github.com/tencentcloudbase/skills --skill auth-tool

If you want to view a Skill's content online first, you can directly open:#!

cloudbase-platform

Trigger Skills

Currently, Skills are not easily activated stably, and it depends on the model and tool. However, engineering methods can improve activation: Skills adopt a progressive loading mechanism: AI first determines relevance based on the Skill's name and description, then loads the complete instructions when needed, avoiding context overflow.

If you want AI to use CloudBase-related Skills more stably, you can:#!

First-line Injection: Add a sentence at the beginning of the question, for example: "You MUST identify the CloudBase scenario first, then read auth-tool / http-api / miniprogram-development / ui-design as appropriate before implementation."
Project-level Rules: Write the same routing constraints in CLAUDE.md or AGENTS.md in the project root directory, so you don't need to repeat them in every conversation.
Hooks (Editor Hooks): Force AI to first perform "scenario identification -> first-read -> before-action check" before sending, which has the highest activation rate. For complete thinking and configuration, see CloudBase Skills: Post-development Person's "Job Manual".

It is recommended to directly write high-frequency routing as explicit rules:#!

Login / Registration / Auth configuration: Read auth-tool first, then read the platform implementation Skill.
Native App / Flutter / React Native: Read http-api first, don't go to Web SDK first.
Mini Program + CloudBase: Read miniprogram-development first, then look at auth-wechat or database Skill.
Pages / Components / Styles: Read ui-design first, output design specifications first, then write UI code.#!

Activation Rate: Low when no intervention is done; first-line injection or project rules can significantly improve; Hooks are highest (about 80%+)#!

Using with MCP

It is recommended to configure CloudBase MCP at the same time: MCP is responsible for secure connection and tool invocation with the CloudBase environment, while Skills are responsible for development specifications and best practices. When used together, AI can both "have permissions" to operate cloud resources and "know the rules" to write code in a production-grade manner.#!

Connection method (local mode / hosted mode) and configuration steps are described in Getting Started.
After configuring MCP in the IDE, install the above Skills, and you can use both tools and skills in conversations.#!

Method 2: Manual Copy

If npx skills add cannot be used in the current environment (or you want to fully customize the content), you can manually copy the rules or prompts to the project, and then let the AI tool read them through methods like "including files".

Where to Copy, How to Use

Rules-type: Place Markdown rules in locations where the project's AI tool can read them, for example:#!
- Cursor: .cursor/rules/ or any path within the project, reference using @Files in conversation
- Claude Code / Others: CLAUDE.md, AGENTS.md, or prompts/*.md
Prompts-type: Save individual prompts as prompts/[name].md, reference them in conversation through the tool's "include file" or # filename, @file and other methods.#!

Usage varies slightly across different tools:#!

GitHub Copilot: #prompts/auth-web.md
Cursor: Select file after @Files
Zed: Select file after /file#!

Available CloudBase prompt scenarios are shown in the "Available Skill Scenarios" below; the complete list of rules and skill packages is available in the CloudBase AI Toolkit repository.#!

Activation Rate: Depends on whether you actively reference these files in conversation. Without active @ or # reference, the probability of AI spontaneously reading them is similar to Method 1 without intervention, and overall it is also low; after active reference, it takes effect on demand.#!

Activation Rate Summary

Method	Characteristics	Activation Rate (Approximate)
Method 1: Install Skills	Easy to use, works well with MCP; can be combined with first-line injection, project rules, or Hooks	Low when no intervention (~20%); first-line injection/project rules can improve; Hooks highest (~80%+)
Method 2: Manual Copy	Flexible, can be fully customized; depends on actively referencing files	Takes effect on demand when actively referenced

You can choose one or more methods from Method 1, or combine with Method 2, according to the project's requirements for "specification stability".

Available Skill Scenarios

Method 1: Install Skills (Recommended)​

Install Skills​

Trigger Skills​

Using with MCP​

Method 2: Manual Copy​

Where to Copy, How to Use​

Activation Rate Summary​

Available Skill Scenarios​

身份认证

数据库

后端开发

应用集成

AI

开发工具

Further Reading​