code_skill

Overview

This skill provides data visualization and inspection capabilities for Python workflows. It creates matplotlib/seaborn plots from files or marimo notebooks, and it can inspect DataFrames by printing rows, columns, and dtypes. The skill handles data-loading inference, generates a self-contained script, executes it via the viz_runner helper, and returns artifact paths for further use.

How this skill works

You supply a visualization spec or a request to inspect data plus a concrete data context (absolute file paths, SQL, or marimo notebook info). The skill infers the data-loading code, builds a complete plotting or inspection script with imports inside the script, and executes it through python /Users/rob/.claude/skills/viz/viz_runner.py. The runner saves a PNG and script under /tmp/viz/, writes metadata JSON, and prints a short summary; the skill then returns the ID and paths to the caller.

When to use it

• Create publication-quality charts (line, bar, scatter, heatmap, violin, histogram).
• Quickly inspect a DataFrame: show first N rows, columns, and dtypes (use show/display).
• Render plots from marimo notebooks without modifying the original notebook.
• Regenerate or refine an existing saved plot by ID or description.
• Convert a SQL query, absolute file path, or code snippet into an executed visualization.

Best practices

• Always provide absolute paths for any files or databases — scripts run from /tmp/viz/ and relative paths will fail.
• Be explicit: include chart type, axes, title, and any special features (reference lines, forecast boundaries, annotations).
• For inspection requests, use language like “show”, “display”, or “first N rows” so show mode (no plot code) is used.
• Provide a suggested ID (e.g., pop_bar) to control naming; the runner will ensure uniqueness by appending suffixes if needed.
• Ask for explicit PNG analysis only when you want the skill to read and interpret the image; otherwise the runner saves PNGs but the skill won’t auto-load them.

Example use cases

• Bar chart of initial vs final members by month from an absolute DuckDB path with a dashed forecast boundary.
• Show the first 10 rows and column dtypes of a CSV at /full/path/to/data.csv (inspection mode using --show).
• Generate a KDE + histogram of a numeric column from a parquet file; use seaborn for distribution features.
• Extract df_forecast from a marimo notebook and plot total_final_members over time (uses --marimo and --target-var).
• Regenerate an existing plot by running the saved /tmp/viz/<id>.py script or by using the runner with the same ID.