Agent Skills: Competition Workflows

Competition Workflows

Overview

Use this skill for Kaggle competition workflows from data access through notebook execution and submission. Prioritize reproducibility, rule compliance, and leakage prevention over leaderboard-chasing shortcuts.

Submission from Notebook Output

Use the notebook-kernel submission form when submitting an output file produced by a notebook version:

kaggle competitions submit <competition> -k <owner/slug> -f <file> -v <version> -m "<message>"

Before submission, confirm the notebook run completed successfully, the output file exists, and the version number matches the intended run. Use notebook-lifecycle for status, logs, files, and output download commands.

Competition Data Access

Use Kaggle CLI or kagglehub for competition downloads. Attach competition sources in kernel-metadata.json when the notebook must run on Kaggle. Confirm users accepted competition rules before assuming downloads or submissions will work.

Submission Validation Checklist

• Assert required columns, order, row count, and ID coverage.
• Check for nulls, infinities, invalid labels, and out-of-range predictions.
• Save outputs under /kaggle/working during Kaggle runs.
• Verify the submitted filename exactly matches the generated artifact.
• Include a concise submission message that identifies model/run changes without exposing secrets.

Leakage and Validation Review

Use group/time/stratified splits that match the competition structure. Apply preprocessing inside folds to avoid fitting transforms on validation data. Avoid using public leaderboard feedback as a validation set; repeated leaderboard probing can overfit. Confirm external data, pretrained models, internet access, and ensemble sources comply with rules.

Reproducibility Standards

Notebook should execute top-to-bottom from a clean Kaggle session. Set seeds for Python, NumPy, framework libraries, and splitters where applicable. Pin package versions when environment drift could affect results. Use a DEBUG flag to run small samples locally or during fast checks without changing final-run logic.

Common Failure Modes

| Symptom | Check | |---|---| | Submission rejected | Filename, columns, rows, competition slug, accepted rules | | Score impossible | Leakage, target contamination, ID mismatch, train/test merge mistake | | Notebook output missing | Save location, run failure, timeout, file pattern | | Local score diverges | Split mismatch, fold leakage, random seeds, preprocessing outside folds |

Safety and Limits

Do not bypass competition rules or suggest hidden test reconstruction, private data scraping, or external data not allowed by rules. Warn before long GPU/TPU training runs. If secrets are needed, advise Kaggle Secrets through the UI; do not claim public API support for secrets administration.

Sources

• Kaggle API docs: https://www.kaggle.com/docs/api
• Kaggle notebooks docs: https://www.kaggle.com/docs/notebooks
• Kaggle kernels CLI: https://github.com/Kaggle/kaggle-api/blob/main/docs/kernels.md