Advanced Applications of Agents in the Skills Era

Advanced Applications of Agents in the Skills Era

Zirui Sheng March 19, 2025 ai agent tutorial

Abstract

This document covers three main questions:

  • What exactly are Skills?
  • How to install, create, and use Skills?
  • What are some recommended Skills and resource libraries?

1. What are Skills?

Skills package instructions, trigger conditions, reference materials, and script resources for high-frequency tasks into reusable capability modules, enabling Agents to complete tasks with less context and greater stability.

Reference video: https://www.bilibili.com/video/BV162cPzhEGU/

To understand Skills, let's look at the entire Agent usage chain:

  • Prompt: Direct instructions to AI
  • Structured Prompt: More complete specification of task objectives, constraints, and output format
  • Command: Sedimenting repeatedly used prompts into fixed instruction files
  • System Prompt: Setting long-term behavioral rules for a project or environment
  • Metadata: Using brief descriptions to help the model quickly determine which files to read
  • References / Scripts: Separating detailed instructions and executable scripts for on-demand loading
  • Skill: Organizing all the above into a complete, reusable capability unit

Skills are not "longer prompts" but "structured task encapsulation". Their goal is not to stuff all knowledge into context at once, but to reduce token consumption through metadata and on-demand loading mechanisms. A useful Skill should tell the Agent both "when to use it" and "how to do it specifically".

In a typical Skills structure:

  • SKILL.md declares what this capability is, what tasks it applies to, and what steps to follow during execution
  • references/ contains detailed knowledge and documentation
  • scripts/ contains directly callable code or scripts
  • assets/ contains templates, icons, styles, sample materials, and other output resources

2. Installation, Creation, and Usage of Skills

1. Different Tools Supporting Skills Have Different Paths

Tool Skills Documentation Global (User-level) Path Project-level Path
Claude Code code.claude.com/docs/en/skills ~/.claude/skills/ .claude/skills/
Codex developers.openai.com/codex/skills ~/.codex/skills/ .codex/skills/
Gemini CLI geminicli.com/docs/cli/skills ~/.gemini/skills/ .gemini/skills/
Opencode opencode.ai/docs/skills ~/.config/opencode/skill/ or ~/.claude/skills/ .opencode/skill/ or .claude/skills/
AMP ampcode.com/news/agent-skills ~/.config/agents/skills/ or ~/.claude/skills/ .agents/skills/ or ~/.claude/skills/
CodyBuddy copilot.tencent.com/docs/cli/skills ~/.codebuddy/skills/ .codebuddy/skills/
Antigravity geminicli.com/docs/cli/skills ~/.gemini/antigravity/skills/ .agent/skills/
VS Code code.visualstudio.com ~/.copilot/skills/ or ~/.claude/skills/ .github/skills/ or .claude/skills/
Cursor cursor.com/cn/docs/context/skills ~/.cursor/skills/ or ~/.claude/skills/ .cursor/skills/ or .claude/skills/
Qwen Code Qwen Code: AI Coding Agent Documentation ~/.qwen/skills/ .qwen/skills
Qoder docs.qoder.com/cli/Skills ~/.qoder/skills/ .qoder/skills
Trae trae.ai/blog/trae_tutorial_0115 ~/.trae/skills/ .trae/skills/
Windsurf docs.windsurf.com ~/.codeium/windsurf/skills/ .windsurf/skills/
Kilo kilo.ai/docs/agent-behavior/skills ~/.kilocode/skills/ ~/.kilocode/skills/
Factory docs.factory.ai/cli/configuration/skills ~/.factory/skills/ .factory/skills/
Goose block.github.io/goose ~/.config/goose/skills/ or ~/.config/agent/skills/ or ~/.claude/skills/ .goose/skills/ or .agent/skills/ or .claude/skills/

2. Three Ways to Install Skills

Method 1: Let AI Help You Install Directly

This is the easiest method. You just need to tell the Agent:

Help me install {skill-name}, the skill's project address is: {skill project address}

This is suitable when:

  • You already know which Skill to install
  • You don't want to manually find directories and handle file hierarchies
  • The current Agent itself has sufficient file operation capabilities

Method 2: Install via Plugin Marketplace

Taking Claude Code as an example:

  1. Enter /plugin
  2. Add the official marketplace: anthropics/claude-plugins-official
  3. Find the target skill in the marketplace
  4. Select installation scope and complete installation

There are 3 installation scopes:

  • User scope: Globally effective
  • Project scope: Effective for all users of the current project
  • Local scope: Only effective for your personal environment in the current project

Method 3: Manual Installation

This is the most universal method, especially suitable for IDE and CLI tools.

  1. Find the skills directory specified by the tool
  2. Create the corresponding folder
  3. Drag the entire directory of the target skill into it
  4. Maintain correct directory hierarchy

If a niche Skill only provides a version for one Agent and you want to use it for another, you can first put it in the directory, then ask the current Agent to help you rewrite it into an adapted version.

3. How to Create Skills

You can let Agents help you create Skills directly:

  1. First, roughly organize existing documentation, commonly used scripts, and reference materials
  2. Then let the Agent help you complete the structure, polish the descriptions, and transform it into Skill format

4. How to Use Skills

  • Agent automatically loads on demand
  • User explicitly specifies which Skill to use

3. Recommended Skills and Resource Libraries

1. Superpower

https://github.com/obra/superpowers

It's essentially a mature software development workflow. Its value lies not just in "brainstorming" but in modularizing the entire process from idea to delivery.

Its workflow includes:

  • brainstorming: Transforming vague ideas into structured designs
  • using-git-worktrees: Establishing independent workspaces for different features to reduce interference
  • writing-plans: Breaking work into clear small tasks with file paths, code, and verification steps
  • Execution: Can follow subagent-driven-development or executing-plans routes
  • test-driven-development: Adhering to the RED-GREEN-REFACTOR TDD cycle
  • requesting-code-review: Code review by severity level
  • finishing-a-development-branch: Verification, merging, or cleanup after task completion

2. planning-with-files

https://github.com/OthmanAdi/planning-with-files

The core idea of this Skill is using the file system as AI's external working memory to alleviate limited context windows and goal drift issues.

It mainly uses 3 Markdown files to manage tasks:

  • task_plan.md: Records phases and progress
  • notes.md: Records research findings and process information
  • [deliverable].md: Consolidates final outputs

3. find-skills

This is "a Skill for finding Skills". Its value is straightforward: when you know you need a certain capability but don't know where existing Skills are, you can use it for retrieval first.

4. research-paper-writing

https://github.com/Master-cai/Research-Paper-Writing-Skills/tree/main/research-paper-writing

This is a Skill for academic writing.

5. Skills Resource Libraries

Claude Code Official Skills Library

Claude Code's official Skills library has two versions, released at different times, with some overlapping Skills:

Four highly-starred awesome-claude-skills repositories on GitHub:

Skillsmp

This Skills website indexes over 80,000 open-source Skills on GitHub, making it the website with the most Skills indexed to date. It supports AI semantic search and keyword filtering, as well as browsing by category and sorting by popularity.

πŸ”—: https://skillsmp.com

Skillsmp screenshot

Skillsmp screenshot

skills.sh: https://skills.sh/

← Back to all posts