在线 Kaggle Notebook 开发工作流。本地编写代码 → push 到 Kaggle 运行 → 拉取结果。 避免大数据集下载，利用 Kaggle 免费 GPU/TPU。 触发关键词：notebook、在线跑、push kernel、远程训练、kaggle run、GPU训练。
Research and analyze a Kaggle competition. Scrapes competition page, reads rules/data description, analyzes top solutions and discussion threads, produces a structured research report. Seeds the IdeaPool with actionable improvement ideas. Trigger: "调研",…
Automated EDA and feature engineering. Generates profiling reports, suggests and creates features based on data patterns and competition type. Includes leakage prevention guardrails and feature quality validation. Trigger: "EDA", "explore data", "数据探索",…
Analyze current performance and drive iterative improvement using tree-search. Uses IdeaPool, ErrorAnalysis, and CodeGen to systematically explore the solution space. NEVER submits during iteration — only generates candidates for review. Trigger: "iterate",…
Train models for the current competition. Handles model selection, cross-validation, hyperparameter tuning via Optuna, and ensemble building. Includes hard-won lessons on leakage prevention, test alignment, and CV reliability. Trigger: "train", "训练模型", "tune…
Meta-skill: automatically upgrade other skills based on new learnings from competition runs. Tracks what worked/failed, updates Lesson Logs, adds new Hard Rules, refines strategies. Trigger: "升级技能", "update skills", "evolve", "what did we learn", "反思",…
Smart submission management with strict daily budget. Never waste submission opportunities. Queue candidates, submit only the best, track CV-LB gap, analyze score trends. Trigger: "submit", "提交", "check leaderboard", "排名", "track score", "submission history",…
AI Agent-driven Kaggle competition workflow. Learn from real competition experience: score stabilization patterns, submission troubleshooting, kernel workflows, GPU task delegation, and the spec-driven development approach that achieved top leaderboard…