kraggle competition

Kaggle competitions are often lost to leaky validation, leaderboard chasing, and unorganized experimentation. Kraggle-competition replaces guesswork with a deterministic Python CLI and a massive 21,980-agent strategy ocean. It enforces a strict "Nautilus" protocol where every notebook must objectively beat the last or face rejection.

What it does

• Mission deployment - Generates a complete battle plan including CV protocols, winner writeup mining, and an ICE-ranked idea queue from a single competition slug.
• Validation locking - Establishes a tamper-evident cross-validation law to prevent public leaderboard chasing and ensure shakeup safety.
• Idea orchestration - Evaluates hypotheses through a "Verdict Council" scoring six axes including feasibility, novelty, and compute cost.
• Monotonic improvement - Tracks every experiment in a "Ratchet" ledger, enforcing a strict one-mutation-per-chamber rule for clear signals.
• Swarm intelligence - Utilizes a mined archive of over 700 competitions and winner writeups to suggest domain-specific techniques and feature stacks.

How it works

1. Initialize - Run kraggle setup <slug> to scaffold the mission workspace and generate the MISSION.md battle plan.
2. Execute - Follow the ranked idea_queue.json, implementing one hypothesis per experiment as dictated by the reverse engineering brief.
3. Evolve - Run kraggle evolve after every notebook execution to log metrics and mutations into the permanent ledger.
4. Endgame - Follow the automated freeze gates for features, models, and ensembling to prepare the final submission duo.

Frameworks & tools

This skill requires Python 3.9+ and uses typer, rich, and requests. It integrates with the Kaggle API for live intel but functions fully offline using a local knowledge graph of techniques and competition history.

Why this beats prompting it yourself

Generic prompts cannot maintain the discipline required for a 3-month competition. This skill provides a deterministic CLI that prevents "silent retries" and forces you to adhere to Gold-medal winning protocols mined from real Grandmaster writeups.

Use cases

• Starting a contest - Deploying a structured war room for a new Kaggle slug within seconds.
• Breaking plateaus - Running the "Fish Tank" swarm to generate and attack new ideas when CV stops improving.
• Submission selection - Applying the "final-two" doctrine to choose the best-CV and best-diversity notebooks.

Known limitations

Requires a manual local install via pip install -e . to enable the CLI. Live data features require a valid KAGGLE_API_TOKEN environment variable.