Skip to main content
The relational-prediction skill guides a coding agent from raw tables and a plain-language question to a scored TabPFN-Rel task. It works in Claude Code and Codex, and installs as a plugin from the TabPFN-Rel repository. With the skill, the agent:
  1. Profiles every table and column, and identifies keys, time columns and columns that leak the future.
  2. Proposes the data preparation and waits for your approval.
  3. Writes the database YAML file and a rerunnable prepare.py.
  4. Proposes two or three prediction tasks and lets you choose.
  5. Writes the task YAML file and checks the label counts and class balance per split.
  6. Compares TabPFN-Rel with constant and LightGBM baselines on held-out data.
The agent asks you before it changes data or picks the task.

​
Install

In a Claude Code session, run:
Start a new session. The skill is listed as tabpfn-rel:relational-prediction.
The skill runs in your Python environment, which needs Python 3.11 or 3.12. On first use, the agent installs tabpfn-rel[api] and the RelArena baselines there. For hosted inference, set TABPFN_TOKEN or run tabpfn_client.init() once; get a token at platform.priorlabs.ai. Hosted fits consume API quota.

​
Use

Open the agent in a directory with your tables (CSV or Parquet) and ask a question about them, for example:
Which of these sellers will receive no orders in the next 30 days?
The agent writes its files to data/<name>/ unless you choose another directory:
To try the skill on a public dataset, point the agent at the Olist database from the TabPFN-Rel cookbook.
Review the preparation the agent proposes. TabPFN-Rel assumes static rows: every value in a row was written at the row’s timestamp and not changed afterwards. Drop any column that violates this, such as an order status or delivery date filled in after the order was placed.