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

Workflow Guide

Every data task in MAGIC follows a five-phase pipeline that moves from raw inputs to finished outputs. The magic-data-lifecycle skill orchestrates this pipeline, routing work to the right specialist skill at each phase.

You don't invoke phases manually. Describe your goal in plain language and the lifecycle skill determines which phase to activate and when to transition.

The Pipeline at a Glance

PhaseSkillWhat Happens
Discovermagic-data-loadingFormat detection, schema inference, initial load
Planmagic-data-lifecycleQuality gating, skill routing, checkpoint strategy
ExecuteDomain skillsCleaning, transformation, analysis, synthesis
Validatemagic-data-validationSchema enforcement, constraint checking, quality gates
Delivermagic-report-generation, magic-data-visualizationReports, charts, exports

Phase Walkthrough

Discover

Skill: magic-data-loading

The agent inspects the input source — a file path, database connection string, or API endpoint — and determines how to read it. Format detection handles CSV, JSON, Parquet, Excel, and more. For databases the agent infers the connection driver and previews the schema.

At the end of Discover the raw data is checkpointed as ckpt_01_raw_data.\{ext\} so that no subsequent phase needs to re-fetch the source.

Key decisions made here:

  • Encoding and delimiter (for delimited files)
  • Row count and memory footprint
  • Whether a profiling pass is warranted before proceeding

Plan

Skill: magic-data-lifecycle (orchestrator)

The lifecycle skill reviews the loaded data and the user's stated goal, then builds a skill route — an ordered list of specialist skills to activate. It also decides:

  • Which quality gates to enforce (e.g. minimum completeness score before transformation)
  • How many checkpoint steps to create
  • Whether the task needs magic-data-profiling before cleaning or can skip straight to transformation

The plan is surfaced as a brief summary in the agent's response. You can ask the agent to explain or adjust the plan before it begins executing.

Execute

Skills: magic-data-profiling, magic-data-cleaning, magic-data-exploration, magic-statistical-analysis, magic-data-transformation, magic-data-synthesis

This is where the actual data work happens. The lifecycle skill activates each specialist in sequence, passing checkpointed outputs between them. Each specialist:

  1. Reads the previous checkpoint
  2. Performs its operation (imputation, normalization, reshape, etc.)
  3. Writes a new checkpoint with a descriptive name

If a step produces unexpected results — for example, a cleaning pass drops more rows than expected — the agent pauses, explains what happened, and asks whether to continue or adjust parameters.

Validate

Skill: magic-data-validation

After execution, the agent runs a validation pass against the final dataset. Validation checks:

  • Schema conformance — column names, data types, and cardinality match the expected schema
  • Constraint checks — nullability, range bounds, referential integrity
  • Quality score — the profiling score must meet the configured threshold (default 85/100)

If validation fails, the agent does not proceed to Deliver. It surfaces the failing checks and recommends targeted fixes. You can ask it to apply fixes automatically or resolve them manually.

Deliver

Skills: magic-report-generation, magic-data-visualization

The final phase produces human-readable outputs from the validated data. Depending on your goal the agent may generate:

  • A Markdown quality report with findings and recommendations
  • Static charts (PNG/SVG via matplotlib) or interactive charts (HTML via Plotly)
  • A cleaned export file in the requested format

All outputs are written under workspace/ so they are easy to find and share.


Skill-to-Phase Mapping


Skipping Phases

Not every task requires all five phases. The lifecycle skill skips phases that are not relevant:

  • No validation needed — if you are exploring raw data without a target schema, the Validate phase is omitted
  • No delivery needed — if your goal is a cleaned file rather than a report, the Deliver phase writes the export and stops
  • No planning needed — single-skill tasks (e.g. "profile this CSV") bypass the lifecycle skill entirely and invoke the specialist directly

Single-skill invocations still save checkpoints and update the workspace state file so the session remains resumable.

Was this page helpful?
Edit on GitHub

Last updated on

On this page