Combining Skills

Individual skills are powerful in isolation, but MAGIC's real strength emerges when skills are chained together. The magic-data-lifecycle skill acts as a router — it inspects the task, selects the right specialist skills, and sequences them into a coherent pipeline.

How Routing Works

When you describe a task, the lifecycle skill:

1. Classifies the goal — a cleaning task, an analysis task, a reporting task, or a combination
2. Identifies required skills — builds an ordered list based on the goal and any quality gates
3. Checks for existing checkpoints — if prior work exists in the workspace, it starts from the most recent checkpoint rather than the beginning
4. Executes each skill in sequence — passing checkpointed outputs between them
5. Applies quality gates — halts if a validation score falls below threshold and asks for guidance

Common Skill Combinations

Load → Profile

The most common starting point. Load raw data and immediately understand its quality before doing anything else.

Triggered by: "Load this file and tell me about it" / "What does this data look like?"

magic-data-loading  →  magic-data-profiling

The loading skill detects format and reads the file. The profiling skill scores quality, identifies nulls and outliers, and reports distributions. Output: ckpt_01_raw_data.csv, quality report.

Load → Profile → Clean → Validate

The standard quality improvement pipeline. Used when you receive a raw dataset and need to prepare it for analysis or hand-off.

Triggered by: "Clean this dataset" / "Prepare this data for analysis"

magic-data-loading  →  magic-data-profiling  →  magic-data-cleaning  →  magic-data-validation

Each step hands off a checkpoint to the next. Validation at the end ensures the cleaned data meets quality thresholds before you use it downstream.

Clean → Validate

Skip loading and profiling when you already have a checkpoint from a previous session.

Triggered by: "Validate the cleaned data from yesterday"

magic-data-cleaning  →  magic-data-validation

The lifecycle skill reads the workspace state file, finds the most recent checkpoint, and starts cleaning from there.

Transform → Visualize

Reshape data and immediately chart the results.

Triggered by: "Pivot the sales data by region and show me a bar chart"

magic-data-transformation  →  magic-data-visualization

The transformation skill produces the reshaped DataFrame and a checkpoint. The visualization skill reads the checkpoint and generates the requested chart type.

Explore → Statistical Analysis → Report

For deeper analytical work: discover patterns, run hypothesis tests, then document findings.

Triggered by: "Analyze revenue trends and give me a report"

magic-data-exploration  →  magic-statistical-analysis  →  magic-report-generation

Exploration identifies segments and patterns. Statistical analysis tests hypotheses and computes correlations. Report generation assembles findings into a structured Markdown document.

Full Pipeline Example: CSV to Report

Here is a complete example — loading a raw CSV, cleaning it, analyzing it, and producing a report.

User prompt: "Load customer_orders.csv, clean it up, and give me a sales analysis report"

Route selected by lifecycle skill:

magic-data-loading
  → magic-data-profiling
  → magic-data-cleaning
  → magic-data-validation       ← quality gate (score ≥ 85)
  → magic-data-exploration
  → magic-statistical-analysis
  → magic-report-generation

What gets created:

workspace/data/checkpoints/
├── ckpt_01_raw_data.csv
├── ckpt_02_profiled.csv
├── ckpt_03_nulls_imputed.csv
├── ckpt_04_duplicates_removed.csv
├── ckpt_05_validated.parquet
└── ckpt_06_analysis_complete.parquet

workspace/data/exports/
├── sales_analysis_report.md
└── revenue_by_region.png

workspace/logs/
└── analysis_journal.md         ← full audit trail

Each step adds an entry to the analysis journal. The final report in exports/ references key findings from exploration and statistical analysis.

Skipping Skills in a Route

The lifecycle skill skips skills that are not needed:

• Profiling is skipped if you explicitly say "skip profiling" or if the data is already profiled (checkpoint exists)
• Validation is skipped if there is no target schema and you have not set quality thresholds
• Exploration is skipped if you only need a transformed export without analysis

You can also insert skills not in the default route by asking: "After cleaning, run a synthesis pass to augment the missing records."

Parallel Skills

Some tasks benefit from running skills in parallel — for example, generating multiple chart types simultaneously. The agent handles this automatically when it detects that two operations are independent and can safely run on the same checkpoint.

Overriding the Route

If the lifecycle skill's chosen route is not what you want, you can correct it mid-pipeline:

• "Skip cleaning and go straight to transformation" — the lifecycle skill adjusts the remaining route
• "Run profiling again on the current checkpoint" — re-runs a skill without advancing the step counter
• "Use magic-data-synthesis to fill in missing rows before validating" — inserts a skill into the route at the current position

All overrides are logged in the analysis journal so the full decision history is preserved.