Introduction

Many users find themselves in a frustrating scenario when using AI tools like Claude or OpenClaw. You start a new conversation to create an Excel spreadsheet and spend three minutes explaining the format, output requirements, and date formats. The AI understands and completes the task well. However, the next day, you return with a new task and repeat the same instructions. This cycle continues daily.

This issue arises not from the AI’s intelligence but from its fundamental mechanism: each new conversation starts as a blank slate, with no memory of previous interactions.

Skill is designed to solve this problem.

Chapter 1: What is Skill?

Imagine starting a new job where your experienced supervisor spends a week training you on various processes and templates. You compile this information into a folder called “Onboarding Manual.” Whenever you face similar situations, you refer to this manual, enhancing your efficiency and ensuring quality.

Skill is like this onboarding manual for AI.

More technically, Skill is a folder containing a document named SKILL.md that instructs the AI on how to handle specific types of tasks, including standards, steps, and output formats. Once created, you place this folder in a designated location, and the AI reads it upon startup. When it encounters matching tasks, it automatically follows the manual without needing repeated explanations.

An AI equipped with Skills functions like a well-trained employee, while one without Skills resembles a new hire who must be retrained daily.

Chapter 2: How is Skill Different from What You Already Know?

Users familiar with AI often ask: Isn’t this just a Prompt? Or how does it differ from MCP?

Both questions are valid and worth discussing.

Difference Between Skill and Prompt

A Prompt is temporary. When you tell the AI in a chat, “You are a professional Excel expert, pay attention to…,” that instruction only applies to that single conversation. Once the chat ends, you must repeat it in a new session.

Skill, on the other hand, is persistent. You write this instruction in a file and place it where the AI can access it. Every time it starts, it knows this Skill exists and automatically uses it for related tasks without needing further input from you.

In simple terms: Prompt is a phone call, Skill is a long-term contract.

Difference Between Skill and MCP

This distinction is a bit more complex. Using a kitchen analogy:

MCP is a professional kitchen equipped with various tools and ingredients, providing the AI with interfaces to access external capabilities, such as connecting to databases or calling APIs.

Skill is like a recipe that tells the AI how to use these tools to create a dish, providing processes, standards, and operational instructions.

Having a kitchen without a recipe means the AI can use various tools but doesn’t know the order to follow for your task. Conversely, having a recipe without a kitchen means the AI understands the process but cannot perform tasks requiring external tools.

They complement each other, with Skill building on MCP to provide structured processes for invoking capabilities.

Another concept often mentioned is Rule. The difference is:

A Rule is like a “company code of conduct” that governs the AI’s tone and format consistently, while Skill is a “job training package” that loads only when relevant tasks arise, guiding the AI on how to perform them professionally.

Chapter 3: How Does Skill Work?

There is a clever design behind Skill that you must understand to create truly useful Skills.

AI does not load all the content of Skills into its memory at once.

Why? Because the AI’s “working memory” is limited, technically referred to as the “context window.” If you load 20 Skills, half the space is occupied before any work begins, overwhelming the AI.

Thus, Skill employs a mechanism called “progressive loading” in three stages:

Stage 1: Discovery

When the AI starts, it only scans the “tags” of each Skill—the name and description—like flipping through a stack of folders and only looking at their labels.

Stage 2: Activation

When your task matches a Skill’s description, the AI reads the complete content of that Skill and begins operating according to the steps outlined. The folder is now opened.

Stage 3: Execution

If the Skill includes scripts or reference documents, the AI only reads them when necessary. Unneeded items remain untouched at the bottom of the folder.

This mechanism directly impacts how you write Skills: the quality of your description determines whether the AI can find and use the Skill at the right moment.

A vague description—“handles various document-related tasks”—leaves the AI unsure when to use it, often leading to it being ignored. A specific description—“use when the user needs to process .xlsx or .csv files for data cleaning, formatting, or chart generation”—enables the AI to quickly determine the appropriate Skill for the task.

Chapter 4: What Does a Skill Look Like?

Don’t be intimidated by technical jargon. A basic Skill is simply a text file in a folder; if you can write, you can create one.

Folder Structure

excel-cleaner/ ← Folder name, choose something meaningful
└── SKILL.md ← Must have, this is the core

A more complex structure may include:

excel-cleaner/
├── SKILL.md ← Must have
├── scripts/ ← Optional: auxiliary scripts (Python, Bash, etc.)
├── references/ ← Optional: reference documents, sample files
└── assets/ ← Optional: templates, resources

As long as there’s a SKILL.md, it constitutes a complete Skill. The rest is optional.

Structure of SKILL.md

This file is divided into two parts.

The first part is the “identity card,” written in YAML format, enclosed between two lines of — . Only two fields are mandatory:

name: excel-cleaner description: Use when the user needs to process Excel or CSV files for data cleaning, formatting, column sorting, deduplication, and chart generation. Not suitable for pure text analysis tasks.

The name is the Skill’s code, which must be all lowercase and can only contain letters, numbers, and hyphens—no spaces.

The description is the Skill’s “trigger condition,” telling the AI when to activate it. This is the most crucial line in the entire Skill.

The second part is the “operation manual,” written in standard Markdown format, detailing how the AI should proceed:

Excel Data Cleaning Skill

Working Principles

• The output file must be in .xlsx format; do not output HTML reports.
• The date column must be unified to the YYYY-MM-DD format.
• Completely duplicate rows should be removed, retaining the first occurrence.
• Numeric columns should not have thousand separators.

Steps

1. Read the file and identify column types.
2. Clean: unify date formats, remove extra spaces, delete duplicate rows.
3. Sort according to user requirements.
4. Output the cleaned .xlsx file, appending _cleaned to the filename.

Example

When the user says, “Clean this CSV, sorted by amount in descending order”:

• Standardize the format.
• Sort by the amount column in descending order.
• Output as original_filename_cleaned.xlsx.

That’s it. It’s that simple.

Chapter 5: How to Install Skills Created by Others

Before creating your own, using existing Skills is the best way to get started.

Installing Skills in Claude Code

Claude Code has its own plugin marketplace, which can be accessed via the command line:

# View available official plugins
claude plugin marketplace

# Install a plugin (e.g., document processing)
claude plugin install plugin_name

Once installed, the Skill will appear in the ~/.claude/skills/ folder. You can also directly copy someone else’s Skill folder into this directory for the same effect. After installation, use the /skills command to list all available Skills and confirm successful installation.

Installing Skills in OpenClaw

OpenClaw features a Skill marketplace called ClawHub (clawhub.ai), similar to an App Store, gathering various community-submitted Skills.

# Search for Skills
clawhub search "automated email sending"

# Install a Skill
clawhub install skill_name

Skills are installed by default in ~/.openclaw/workspace/skills/, becoming effective after restarting OpenClaw.

Installing Skills in Hermes Agent

Hermes has a Skills Hub where Skills are shared within the community. After installation, they are placed in the corresponding directory and automatically scanned and loaded when Hermes starts.

The core logic across these three platforms is the same: find the folder, place it in the right location, restart, and it becomes effective.

Chapter 6: Manually Writing Your Own Skill—Starting from a Real Scenario

With the theory covered, let’s create one.

Scenario: You need to write a weekly work report with a fixed format and style, explaining it every time.

Step 1: Create a folder in an appropriate location on your computer.

If using Claude Code, enter in the terminal:

mkdir -p ~/.claude/skills/weekly-report

If using OpenClaw, replace ~/.claude with ~/.openclaw/workspace, and the rest remains the same.

Step 2: Create a SKILL.md file in this folder using any text editor:

name: weekly-report description: Use when the user says to write a weekly report, generate a report, or summarize this week’s work. Outputs a structured weekly report containing completed tasks, next week’s plan, and encountered issues.

Weekly Report Generation Skill

Output Format

The report must contain three sections, with none omitted:

1. Completed Tasks

• List in concise verb-object structure.
• Each entry should be within 30 words.
• Include key data or results (if any).

2. Next Week’s Plan

• List 3-5 clear key tasks for the next week.
• Clearly state what to do, avoiding vague expressions.

3. Encountered Issues

• If there are unresolved issues, list and explain their impact.
• If there are no issues, write “No significant obstacles this week.”

Writing Style

• Use concise language, avoiding unnecessary words.
• Use first-person or objective statements, avoiding “we.”
• Do not include introductory fluff; directly output the main content.

Output Requirements

• Markdown format.
• Heading levels: use ## instead of #.
• Do not add phrases like “Further details can be added if needed” at the end.

Step 3: Save and restart your AI tool.

Step 4: Next time, simply say, “Help me write this week’s report,” and provide your completed tasks. It will automatically output in the defined format without you needing to specify any formatting requirements.

Chapter 7: Key Points for Writing a Good Skill

After completing the basic version, here are some details that make Skills truly useful.

First, write the description as a “trigger condition,” not a “function description.”

Many first-time writers describe the function, like “This skill helps with Excel-related data processing.” This describes functionality but does not inform the AI when to use it.

Change it to: “Use when the user mentions processing Excel or CSV files, or needs data cleaning, deduplication, or formatting. Not suitable for simple data queries or pure text analysis.” Adding when to use and when not to use significantly improves the AI’s accuracy.

Second, keep SKILL.md to a maximum of 500 lines.

This is not to say longer content is unacceptable, but excessive length consumes more context space, reducing the AI’s ability to handle other tasks. If the content is extensive, place reference documents in the references/ folder and keep the core process in SKILL.md, referencing them as needed.

Third, consider using scripts.

If your task requires executing certain fixed operations, place Python or Shell scripts in the scripts/ folder and instruct the AI in SKILL.md to execute them at specific steps. This guarantees consistent outputs—same input always yields the same output, unlike AI-generated text, which may vary.

Fourth, for high-risk operations, consider manual triggering.

Some Skills may involve irreversible actions like modifying files, sending messages, or deploying code. For these, add the line:

disable-model-invocation: true

This line ensures the Skill can only be triggered manually by entering the Skill name, preventing accidental activation by the AI.

Fifth, test the trigger accuracy after writing.

After installing the Skill, test two scenarios:

1. Should Trigger: Make a request based on the scenarios described to see if the AI uses this Skill.
2. Should Not Trigger: Make an unrelated request to check if the AI mistakenly uses this Skill.

If it fails to trigger when it should, add more potential keywords to the description. If it triggers incorrectly, clarify in the description what it is not suitable for.

Chapter 8: Where to Place Skills and How to Find Them

Using existing resources can save you a lot of trouble.

Official Repository

Anthropic has an official Skills repository on GitHub (github.com/anthropics/skills) containing 18 official Skills, including PDF processing, Excel operations, Word documents, algorithm art, brand guideline applications, code testing, etc. These are reliable and can be used directly or as learning references.

Community Market

ClawHub (clawhub.ai) is the Skill market for the OpenClaw ecosystem, currently aggregating thousands of community-submitted Skills covering automation workflows, document processing, information gathering, and more. You can search and install them directly.

Hermes Skills Hub

Hermes Agent also has its own community for sharing Skills, primarily targeting workflows requiring continuous operation and cross-session memory.

Local Installation Location

Most AI tools store Skills in hidden folders within the user directory:

• Claude Code: ~/.claude/skills/
• OpenClaw: ~/.openclaw/workspace/skills/

You can copy any properly formatted Skill folder into these locations, restart the tool, and it will be available. This is the simplest way to install Skills without needing command-line operations.

Chapter 9: The Complete Lifecycle of a Skill