skills: - ms - agents - mcp --- AI agents like Claude Code are powerful, but they don't know your specific project conventions or unique workflows out of the box. **Skills** are reusable prompt templates that give agents new capabilities. They are like "plugins" for your AI. We use **meta_skill (ms)** to manage these skills locally. It's a local-first knowledge management tool with hybrid semantic search (BM25 + hash embeddings) and Git-backed audit trails. Your skills live in your project re...
Installs into .claude/skills of the current project.
Are you the author of Lessons?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/quangdang46-lessons)
# Lesson 11: Local Skill Management with ms
skills:
- ms
- agents
- mcp
---
# What are Skills?
AI agents like Claude Code are powerful, but they don't know your specific project conventions or unique workflows out of the box.
**Skills** are reusable prompt templates that give agents new capabilities. They are like "plugins" for your AI.
We use **meta_skill (ms)** to manage these skills locally. It's a local-first knowledge management tool with hybrid semantic search (BM25 + hash embeddings) and Git-backed audit trails. Your skills live in your project repo, not in the cloud.
---
# Checking Installation
First, let's make sure `ms` is installed and healthy.
Run the version check:
```bash
ms --version
```
Then run the doctor check to verify everything is configured correctly:
```bash
ms doctor
```
---
# Initializing a Project
To use skills in a project, you initialize `ms`. This creates a `.ms/` directory where your project-specific skills will live.
Try initializing in your current directory:
```bash
ms init
```
You should see a message confirming that `.ms/config.toml` was created.
---
# Discovery & Suggestions
`ms` uses a smart search algorithm (Thompson sampling) to suggest the most effective skills for a given task.
Imagine you want to write a test. Ask `ms` for suggestions:
```bash
ms suggest "write a unit test"
```
If you had skills installed, it would show you the best matching ones.
You can also list all available skills:
```bash
ms list
```
---
# Creating a Skill
Let's create a simple skill. Skills are just Markdown files with some YAML metadata.
Create a new skill called "hello":
```bash
ms create "hello"
```
This will open your editor. You can define what the skill does (e.g., "Say hello to the user in a pirate voice").
---
# The MCP Connection
The magic happens when you connect `ms` to Claude Code via the **Model Context Protocol (MCP)**.
ACFS sets this up automatically! When you run Claude Code (via `cc`), it can "see" your `ms` skills as native tools.
If you ever need to start the MCP server manually (rare), you can run:
```bash
ms mcp serve
```
(Press Ctrl+C to stop it if you run it)
---
# Summary
You've learned:
1. **ms** manages AI skills locally in your project.
2. **ms init** sets up a project.
3. **ms suggest** finds the right skill for the job.
4. **MCP** connects these skills to Claude Code automatically.
Next, we'll look at how to find *remote* skills from the community using `jfp`.