TabPFN Prediction Skill for Claude Code

A Claude Code skill that enables predictions on structured/tabular data using TabPFN via the tabpfn MCP server.

What it does

Adds a /predicting-with-tabpfn skill that lets Claude run classification and regression predictions on tabular data — CSVs, spreadsheets, or inline tables — without writing any ML code.

Trigger phrases: "predict X from this data", "classify these rows", "estimate/score/forecast", or any mention of TabPFN or Prior Labs.

Requirements

• tabpfn MCP server configured. See here how to configure the MCP server.

Installation

Copy .claude/skills/SKILL.md into your project's .claude/skills/ directory (or your global ~/.claude/skills/ for use across all projects).

Usage

Once installed, Claude will automatically invoke the skill when you ask it to make predictions on tabular data. It will confirm the target column, task type, and train/test split before calling any tools.

Supports:

• Classification and regression
• Inline data (small tables pasted into chat)
• File-based data (CSV upload)
• Re-prediction on new data using a saved model ID

Data limits: up to ~50k rows, ~2000 features, and 20M total cells (train + test).

How it works

The skill guides Claude to use the tabpfn-stg MCP tools directly:

• fit_and_predict_inline — for small tables already in the conversation
• upload_dataset + fit_and_predict_from_dataset — for CSV files
• predict_inline / predict_from_dataset — to reuse a trained model

TabPFN handles missing values, categoricals, and unnormalized numerics natively — no preprocessing required.