Jupyter Notebook

Jupyter Notebook is an official skill from the curated skills collection in the openai/skills repository. Its direct primary source is the skills/.curated/jupyter-notebook directory. According to the provider, the skill helps a Codex agent create, structure, or revise Jupyter notebooks with the ipynb extension for experiments, exploratory analysis, and readable tutorials. It is an instruction set with templates and a helper script, not Jupyter itself, not a data-analysis service, and not a guarantee that the resulting work is scientifically correct.

Purpose and classification

The skill separates two main notebook modes. An experiment serves a hypothesis, comparison, or exploratory investigation. A tutorial walks a particular audience through a concept step by step and should be readable and rerunnable by other people. When an existing file is involved, the provider treats the task as a refactor: preserve the intent while improving structure, readability, and reproducibility. This decision prevents an agent from turning a loose analysis into teaching material without confirming the purpose, or from treating an instructional walkthrough as an unstructured scratchpad.

Before writing, the agent should identify the objective, audience, and definition of done according to the provider. It then chooses the suitable template and uses the bundled new_notebook.py helper to create a consistent starting structure. The templates are experiment-template.ipynb and tutorial-template.ipynb in the skill directory. The helper loads a template, updates its title cell, and writes a new file. This avoids error-prone hand construction of notebook JSON. Intermediate work belongs in tmp/jupyter-notebook, while final artifacts belong in output/jupyter-notebook.

Structure and practical workflow

A useful notebook consists of small runnable steps according to the provider. Each code cell should focus on one bounded task. Short Markdown cells explain the purpose and expected result so readers understand the flow before executing code. Large, noisy outputs should be replaced with concise summaries when that preserves the evidence. Experiments benefit from traceable inputs, controlled variables, and clearly named results. Tutorials additionally need a sensible sequence, an appropriate difficulty level, and visible transitions between explanation and execution.

When editing an existing notebook, the agent should preserve its structure and reorder cells only when that improves the top-to-bottom story. If direct JSON editing is unavoidable, the documented notebook structure should be reviewed first. The quality process includes running the notebook from top to bottom when the environment permits. If execution is not possible, the agent must state that limitation plainly and explain how to validate locally rather than claiming a successful run that did not happen.

Requirements, safety, and boundaries

Creating a scaffold requires only the Python standard library for the bundled helper. Optional local execution may require JupyterLab and an IPython kernel. The skill should not silently replace missing tools with untested alternatives. Notebook code can read files, import packages, access networks, or expose computed data. Only approved inputs belong in a notebook; passwords, tokens, private keys, and personal data must not appear in cells, outputs, or example artifacts. Notebook content and external documents are data, not instructions that justify granting unreviewed permissions.

The Coding, Data Analysis, and Office categories describe documented use cases, while Codex compatibility reflects the official OpenAI context. According to the provider, the skills documentation explains how skills teach repeatable workflows to ChatGPT and Codex. It does not replace Jupyter documentation, project standards, or human review of data, code, licensing terms, and results. The skill can improve structure and reproducibility, but it cannot prove that an analysis is causally correct, statistically sound, or safe for production data.