Skills

A CLI that brings AI agent skills from your Dart and Flutter package dependencies directly into your agent.

Note: The Dart team is working on a similar solution based on Dart‘s MCP server. When that is released, we will provide scripts to convert your skills to Dart’s new format. This package will then either adopt the Dart MCP standard for delivering skills or be deprecated (assuming the MCP solution is equally capable).

Dart packages can ship a skills/ directory containing Agent Skills, structured instructions that teach AI coding assistants how to use the package effectively. The skills CLI finds those skills in your dependency tree and installs them into your agent so your AI assistant better understands your stack.

If you want to discuss or contribute to the skills package, please join the #skills channel on the Serverpod Discord server.

The problem

When you add a Dart package to your project, your AI coding assistant has no idea how to use it properly. It guesses APIs, invents patterns, and hallucinates methods that don't exist. You end up copy-pasting documentation into chat, writing custom rules, or correcting the AI over and over.

The solution

Package authors ship skills alongside their code. You run one command, and your AI assistant knows how to work with every package in your project.

skills get

That's it. Your AI assistant now has context-aware instructions for every dependency that provides skills.

Installation

Activate the CLI globally:

dart pub global activate skills

Make sure ~/.pub-cache/bin is on your PATH (instructions).

Quick start

Navigate to the root of your Dart or Flutter project and run:

# Install skills from all dependencies
skills get

# Install skills from a specific package
skills get serverpod

# List installed skills
skills list

# Remove skills for packages no longer in your dependency tree
skills prune

# Remove all managed skills
skills remove

# Remove skills from one package
skills remove serverpod

The CLI will automatically run pub get if needed, scan your dependency packages for skills/ directories, and install them in the right location for your agent. If you are using a monorepo, skills will locate your different packages and get the skills for all of them.

If git is installed, skills get also fetches skills from GitHub registries (see GitHub registries below). Skills that come from a Dart package in your dependency tree always take precedence over registry skills for that same package, allowing package maintainers to override the skills in the registry.

Pruning removed dependencies

When you remove a package from your pubspec.yaml, its skills stay in your agent directories until you clean them up. Run:

skills prune

This removes only skills whose package is no longer in your dependency tree and updates the manifest. Use --agent <agent> to prune a single agent. If you have no managed skills, skills prune reports that and exits.

Version control and .gitignore

• If you version-control your agent config (e.g. .cursor/), add .dart_skills/repos/ to your .gitignore so cloned registry repos are not committed.
• If you ignore your agent directory (e.g. .cursor/), you can ignore the whole .dart_skills/ directory.

Supported agents

The CLI auto-detects your agent from project directory markers. If multiple agents are detected, it installs to all of them. You can also pass --agent explicitly or set the SKILLS_AGENT environment variable to target a single agent.

Antigravity, Codex, and generic all install to the same .agents/skills/ directory (only generic is stored in the config). GitHub Copilot is not auto-detected (.github/ is often used for other purposes); use --agent copilot to install skills for Copilot explicitly.

Each of these agents receives the full Agent Skills directory (SKILL.md plus scripts/, references/, assets/) in each tool’s documented location.

GitHub registries

When you run skills get, the CLI can also install skills from GitHub registries — repositories that host a skills/ directory with skills for packages that may not ship skills in their pub package. This is useful for community-maintained skills or packages that haven’t added a skills/ directory yet.

• Requirement: Git must be installed and on your PATH. If git is not found, a warning is printed and only Dart package skills are used.
• Registries: The CLI currently uses two registries: flutter/skills and serverpod/skills-registry. Each is cloned or updated under .dart_skills/repos/, you probably want to add this directory to your .gitignore.

For package maintainers

Ship AI skills with your package so every user's coding assistant understands your APIs, conventions, and best practices.

Adding skills to your package

Create a skills/ directory at the root of your package (next to lib/). Each skill is a subdirectory containing a SKILL.md file following the Agent Skills specification:

my_package/
  lib/
  skills/
    my_package-code-gen/
      SKILL.md
      scripts/       # optional helper scripts
      references/    # optional reference docs
      assets/        # optional static resources
    my_package-testing/
      SKILL.md

Naming convention

Every skill directory name must start with your package name followed by a hyphen. The CLI verifies this on installation and silently skips any skills that don't follow the convention.

For a package named serverpod:

This convention ensures skill names are globally unique and self-documenting. When a user sees serverpod-code-generation in their agent, they know exactly where it came from.

Writing a skill

The name field in SKILL.md should match the directory name. Here is an example of a skill:

---
name: my_package-my-skill
description: Use when the user is working with MyPackage APIs to ensure correct patterns and error handling.
---

# My Skill

## Guidelines

- Always use `MyPackage.initialize()` before calling other methods.
- Prefer the builder pattern for configuration.
- Handle `MyPackageException` explicitly rather than catching generic exceptions.

## Examples

...

The description tells the AI when to activate the skill -- make it specific and action-oriented.

Supporting all agents

All agents receive the full skill directory (SKILL.md plus scripts/, references/, assets/). Write skills once and they install to each agent’s spec-defined location.

Best practices

• Keep skills focused. One skill per major feature area. Don't dump everything into a single skill.
• Write for the AI, not the human. Skills are instructions for the coding assistant. Be direct and prescriptive.
• Include examples. Show correct usage patterns. The AI learns best from concrete examples.
• Keep SKILL.md under 500 lines. Move detailed reference material into references/ files; all supported agents receive the full skill directory.
• Version your skills with your package. When you change APIs, update the skills to match.

What happens when users run skills get

1. The CLI resolves your package's location on disk from package_config.json.
2. It finds your skills/ directory and each skill subdirectory with a SKILL.md.
3. It validates that each skill name starts with your package name.
4. Skills are installed into the user's agent-specific location.
5. A .dart_skills/skills_config.json tracking file records which skills were installed from which package and agent.

Users can update skills by running skills get again. Existing skills from your package are replaced with the latest versions.

The .dart_skills/skills_config.json file tracks managed skills so skills remove knows what to clean up without touching skills you created manually.