MAGIC Agent Skills is now open source! Star on GitHub
MAGIC Agent SkillsMAGIC Agent Skills
Getting Started

How It Works

MAGIC Data Agent Skills transforms any AI coding assistant into a data agent — an assistant specialized in data engineering, data processing, and data science workflows. It does this through self-contained knowledge packages called skills.

Each skill teaches your AI coding assistant how to think about data tasks through three layers:

The 3-Layer Architecture

Domain Knowledge

What the agent needs to understand about the task domain. Includes natural-language triggers (when to activate), rules, constraints, and anti-patterns to avoid.

Code Patterns

Seed patterns, templates, and reference implementations the agent studies and adapts to your specific task. These are not executed directly — the agent reads them and writes its own custom code tailored to your data.

Procedures

Step-by-step workflows the agent follows. Reference scripts in scripts/ demonstrate approaches — the agent studies them for patterns, then writes code adapted to the specific task at hand.

How the Agent Uses Skills

When your AI coding assistant receives a data task, it follows this process:

  1. Reads SKILL.md — gets domain knowledge, code patterns, and procedures
  2. Reads references — detailed reference docs loaded on demand for deeper context
  3. Reads scripts — studies reference implementations to understand proven approaches
  4. Writes its own code — adapts patterns to your specific data, format, and requirements
  5. Verifies quality — uses the constraints and rules from the skill to validate its work

Skills provide knowledge and patterns. The agent decides how to act — it may follow the reference scripts closely, adapt them significantly, or write entirely custom code depending on the situation.

Agent Autonomy

Skills guide — they don't force. Your AI assistant reads the SKILL.md, understands the domain, and makes intelligent decisions about how to apply the patterns to your specific situation. This means:

  • The agent adapts code to your data format and structure
  • It skips unnecessary steps when they don't apply
  • It combines multiple skills when your task spans domains
  • It explains its reasoning as it works

The Workflow

Every data task follows a natural pipeline orchestrated by the magic-data-lifecycle skill:

DiscoverPlanExecuteValidateDeliver

Each phase routes to the right skill automatically. Not every step is needed for every task — the agent adapts the sequence to what your data actually needs.

Was this page helpful?
Edit on GitHub

Last updated on

On this page