kaggle-skill

Ask your coding agent about a Kaggle competition and get a short, sourced
answer. Let it track where you stand, check a submission, and submit only
after you say yes.
An independent, unofficial skill and plugin named kaggle, for Claude Code,
Codex and other agents that load SKILL.md packages. Not affiliated with,
endorsed by, or sponsored by Kaggle or Google.
> **You:** Which Kaggle competitions with prize money are running right now? A short list with each one's deadline, prize and metric, please.
A real session, recorded 2026-10-02 in Claude Code 2.1.286 (claude-opus-5-5): the agent ran competitions, then answered. The whole answer, cast.
What you need
- Python 3.11 or later, and an agent that loads skills.
- Nothing else to read public content: a competition's facts and pages,
solution writeups, discussions.
- For your own account (status, downloads, submissions, notebooks,
publishing): a Kaggle credential, and two packages for the python3 your
agent uses: python3 -m pip install "kaggle>=2.2.4" "kagglehub>=1.0.2".
Install
1. Add it to your agent.
| Where | Command |
|---|---|
| Claude Code | /plugin marketplace add shepsci/kaggle-skill then /plugin install kaggle@shepsci |
| Codex | codex plugin marketplace add shepsci/kaggle-skill --ref main then codex plugin add kaggle@shepsci |
| skills.sh (Antigravity CLI, Cursor, and others) | npx skills add shepsci/kaggle-skill |
| OpenClaw | clawhub install kaggle |
| By hand | Clone this repository and copy skills/kaggle/ into your agent's skills folder |
2. Start a new session and ask something that mentions Kaggle: "What is the
metric and deadline of the Kaggle Titanic competition?" The skill loads
when Kaggle is named. It pre-approves only reading files, so in Claude
Code's default mode the agent asks before it runs each command.
3. For anything on your account, sign in once with kaggle auth login, or
create a token with "Generate New Token" at
kaggle.com/settings. Then ask "check my
Kaggle setup", which runs the skill's doctor command. The
setup guide has
the details.
In Claude Code, shepsci/kaggle-skill is a marketplace that holds one
plugin, which is why there are two commands.
What you can ask
- "Which Kaggle competitions with prize money are running, and what are their metrics?"
- "What is the metric, the deadline and the prize of this Kaggle competition?"
- "Summarize the rules that matter: team size, external data, submission limits."
- "What did the top teams do? Preview their solution writeups."
- "What are people discussing in this competition this week?"
- "Where do I stand: rank, submissions left today, best score, GPU hours?"
- "How far am I from the bronze line, and what moved since yesterday?"
- "Is this submission file well formed?"
- "Submit it." (You get a dry run first, and it submits after your yes.)
- "Wait for the score and tell me how it compares with my validation."
- "Download this dataset." "Push this notebook and tell me when the run ends."
What it does on its own, and what it asks first
| Action | Without asking | What it costs |
|---|---|---|
| Read competition facts, pages, writeups, discussions | Yes | Nothing |
| Read your rank, submissions, scores, quota | Yes, with your credential | Nothing |
| Download competition data | Yes; refuses above 20 GB unless told | Disk space |
| Check a submission file | Yes | Nothing |
| Submit to a competition | No: a dry run, then --yes after you agree | One of the day's submissions |
| Push or run a notebook | No: a dry run first | Weekly GPU hours when a GPU is on |