topprismdata

@topprismdata source repo

64 skills
0 followers
10 hours ago last updated

▌
Skill Tester · topprismdata bundle
Tests and evaluates any Claude Code skill for structural validity, quality, and trigger accuracy. Implements the cc-plugin-eval 4-stage pipeline (Analysis → Generation → Execution → Evaluation) and the 4D scoring rubric (Documentation/Code/Completeness/Usability 25% each). Use before packaging or deploying any skill.

0
installs
▌
Claudeception · topprismdata
Extract reusable knowledge from work sessions into new Claude Code skills. Use when: (1) /claudeception command to review session, (2) "save this as a skill" or "extract a skill", (3) "what did we learn?", (4) After non-obvious debugging, workarounds, trial-and-error discovery, or counterintuitive solutions. Do NOT trigger for routine fixes or straightforward tasks with obvious solutions.

0
installs
▌
Ml Sweet Spot Principle · topprismdata
ML optimization sweet spot principle: "More is not always better". Use when: (1) Increasing model complexity but validation score plateaus or drops, (2) OOF keeps improving but LB stays same or gets worse, (3) Debating between simpler vs more complex models, (4) Feature selection or hyperparameter tuning, (5) Considering more seeds in ensemble (beyond 5-10), (6) Research claims show unexpected results at your scale, (7) Combining different feature engineering approaches. Covers OOF-LB gap analysis, overfitting detection, ensemble sweet spots, research claim validation, and feature combination incompatibility. **UPDATED 2026-06-14**: Added CatBoost-First evidence (Spaceship Titanic) and AutoGluon-First comparison (House Prices V18 0.1194 vs AutoGluon 0.1180). 5-variant CatBoost ensemble is the sweet spot for manual GBDT work. AutoGluon 5-15 min should be the first step, not the last.

0
installs
▌
Model Ensemble Negative Weight Effect · topprismdata
Avoid ensembling models with significantly different performance. Use when: (1) Considering weighted average of models with divergent OOF/LB scores, (2) One model clearly underperforms another, (3) Ensemble shows lower validation score than best individual model. Covers the negative weight effect where adding a weaker model reduces overall performance despite theoretically sound weighting strategy.

0
installs
▌
Autogluon First · topprismdata
AutoGluon-First Strategy: Always run AutoGluon `best_quality` preset as the first step in any tabular ML competition (5-15 min baseline). Validated 3/4 times vs manual ensembles on small/medium tabular datasets. Use when: (1) Starting any new tabular competition, (2) Need a strong baseline in <15 minutes, (3) Manual GBDT work takes hours but gives similar results, (4) Want to focus manual effort on complementary techniques (deep learning, external data, special features) rather than reimplementing what AutoGluon already does automatically. Covers 5 core reasons for success (multi-algorithm diversity, multi-level stacking, automated preprocessing, bagging/multi-fold, Bayesian optimization), when to AVOID AutoGluon (text/image data, multi-output regression, >1M rows with limited RAM), and how to use it as a Silver signal in custom pipelines.

0
installs
▌
Multi Agent Roles · topprismdata
Use when designing a multi-agent system, when a single agent has too many responsibilities, when planning complex ML pipelines, or when observing "agent attention fragmentation" symptoms. Triggers when a task involves planning + coding + review steps.

0
installs
▌
Runtime Reflexion · topprismdata
Use when running Python code via sandbox, when an experiment fails with a runtime error, when building self-correcting ML agents, or when debugging auto-generated code. Triggers on subprocess failures, ImportError, CUDA OOM, shape mismatch, or any "agent gets stuck on syntax error" pattern.

0
installs
▌
Arxiv Paper Search · topprismdata
Use when encountering a new ML technique not covered by existing 43+ skills, when needing citations for a paper/report, when comparing recent (last 6 months) approaches to a known problem, or when the existing knowledge feels stale. Triggers on phrases like "latest paper on X", "SOTA in Y", "what does the literature say about Z".

0
installs
▌
Context Engineering · topprismdata
Use when context window is getting crowded, when dealing with long training logs, or when preparing prompts with mixed-priority content (system / task / skills / examples). Triggers when assembling prompts >4000 chars, or when observing "Lost in the Middle" symptoms (model ignores mid-prompt content).

0
installs
▌
Yoy 364day Features · topprismdata
Use when: (1) building time series features for data with annual seasonality, (2) working with retail/sales forecasting where year-over-year patterns matter, (3) looking for features that capture long-term seasonal patterns beyond short lags, (4) the 1st place solution mentions "364-day" or "YoY" features. NOT for: data without annual seasonality, very short history (< 1 year).

0
installs
▌
Gsd Loop Engineering · topprismdata
Apply GSD Core's loop-engineering methodology to ML/data-science tasks. Use when: (1) You have a multi-step ML pipeline (data → features → train → verify → submit), (2) Context is growing long and quality is drifting, (3) You want auditable verification at each step, (4) You want fresh-context subagents to handle heavy work without polluting the main session. Validated against: gsd-build/get-shit-done (64K stars, deprecated), open-gsd/gsd-core (canonical, 4.4K stars, active 2026-06).

0
installs
▌
Mcts Pipeline Search · topprismdata
Use when exploring hyperparameter combinations for ML models, when doing Top-1% push on Kaggle, when grid search is too expensive, or when needing systematic exploration of pipeline variants. Triggers for neural architecture search, ensemble weight tuning, or feature subset selection.

0
installs
▌
Retail Eda Framework · topprismdata bundle
Comprehensive EDA approach for retail/fashion/tabular ML using best-in-class libraries. Use during stage 1 (data understanding) of any ML pipeline. Built around 5-stage pipeline: (1) data quality with ydata-profiling + missingno, (2) statistical profiling with sweetviz, (3) domain-specific (RFM, transaction patterns, co-occurrence), (4) time-series with tslumen, (5) summary. Validated on H&M Personalized Fashion Recommendations (105K articles, 25 cols, 0.39% missing — extremely clean).

0
installs
▌
Self Evolving Skills · topprismdata
Use when adding new skills to the library, when extracting insights from completed experiments, when consolidating lessons learned, or when running a project retrospective. Triggers after a successful Kaggle competition, after major model breakthrough, or when "we should remember this" comes up.

0
installs
▌
Kaggle Experiment Sop · topprismdata
Standard Operating Procedure for ANY Kaggle experiment or competition entry. Execute sequentially from Phase 0 to Phase 3. Each phase has explicit checklist items with pass/fail criteria. Covers: memory recall, competition type identification, public kernel evaluation, baseline establishment, controlled iteration, OOF/LB validation, submission format verification, and post-experiment knowledge crystallization. This SOP synthesizes 4+ months and 20+ competitions of experience into a single repeatable workflow.

0
installs
▌
Catboost First Tabular · topprismdata
CatBoost-First Strategy: When manual GBDT work is needed, start with CatBoost (not LightGBM or XGBoost). CatBoost has native categorical feature handling, robust to overfitting, and consistently outperforms other GBDTs on small/medium tabular datasets. Use when: (1) AutoGluon is not available or too slow, (2) Need to add CatBoost as a Silver signal in custom pipeline, (3) Manual GBDT baseline required, (4) Categorical features are dominant (more than 5-10 high-cardinality columns). Validated: Spaceship Titanic 0.8124 OOF acc (vs XGBoost 0.8003, LightGBM 0.8048), House Prices V17 CatBoost alone stronger than XGBoost. Covers native categorical handling, ordered boosting, robust default parameters, ensemble sweet spot (5 CatBoost variants > 10+ same-family variants).

0
installs
▌
Agent Nurture Framework · topprismdata
Agent Nurture Framework: a systematic methodology for training AI agents from novice to expert through conversational knowledge crystallization. Use when: (1) designing an agent training pipeline, (2) consolidating fragmented skills into organized knowledge architecture, (3) integrating external learning resources (books, notebooks) into agent memory, (4) measuring agent capability growth over time. Based on 2 months of real-world experimentation with 156 skills and 51 memory documents across multiple ML competitions and projects.

0
installs
▌
Kaggle Optimal Blending · topprismdata
Optimal blending strategy for Kaggle competitions: 80/20 rule (re-ranking vs raw). Use when: (1) Re-ranking gives unexpected results, (2) Simple average blending underperforms, (3) Single model dominates ensemble. Avoids the trap of 100% re-ranking or equal-weight ensembles. **UPDATED 2026-06-14**: Added asymmetric-blending principle (30/70 rule) for general tabular ensembles. When one model family dominates (e.g., CatBoost in Spaceship Titanic 0.8124 OOF acc), weight it higher than other families. Validated: 30% Silver + 70% Top-5 = 0.8132 OOF (best on SST V3).

0
installs
▌
Okf Visualize Knowledge · topprismdata
Use Google's Open Knowledge Format (OKF) to convert any directory of markdown files into a force-directed knowledge graph. Validated on `docs/ml-agent-memory/`: 12 concepts / 35 edges rendered in 5 seconds. Use when: (1) You have a directory of markdown notes with cross-references, (2) You want to visualize how your concepts/skills/lessons connect, (3) You want to add type-driven coloring (Principle vs Skill vs Lesson vs Competition vs Bundle), (4) You want self-contained HTML output that runs offline.

0
installs
▌
Cv Lb Gap Acknowledgment · topprismdata
CV-LB Gap Acknowledgment: CV improvement does NOT equal LB improvement. This is one of the most important MLOps principles. Use when: (1) OOF score keeps improving but LB score plateaus or drops, (2) Spending days tuning hyperparameters without LB improvement, (3) Comparing models on CV only without LB validation, (4) Trusting cross-validation as the "final" metric. Covers the mathematical reasons for CV-LB gap (overfitting to OOF noise, distribution shift train vs test, hyperparameter over-tuning to CV), empirical validation (0.005-0.01 gap observed in tabular competitions), the 5-stage validation pipeline (CV → submission → LB → analysis), when to stop iterating on CV alone, and the 80/20 rule for time allocation between CV optimization and LB validation.

0
installs
▌
Gpu Readiness Assessment · topprismdata
Use when: (1) Considering switching from CPU to GPU for a Kaggle competition, (2) GPU quota is running low and you need to prioritize which competitions deserve GPU time, (3) About to start a model training run and unsure if GPU is worth the 30h/week quota cost, (4) Deciding between CPU AutoGluon vs GPU neural network. Provides a 5-gate assessment framework that produces a verdict: READY (GPU will help significantly), MARGINAL (GPU helps but ROI is low), or SKIP (CPU is sufficient). Inspired by NVIDIA's cupynumeric-migration-readiness, adapted for Kaggle GPU quota management.

0
installs
▌
Kaggle Data Format First · topprismdata
Prevent wasted research by verifying Kaggle competition data format BEFORE investing in RAG, technical planning, or model architecture design. Use when: (1) Starting any new Kaggle competition, (2) Competition name/size is ambiguous about data format, (3) Planning to do extensive research before implementation, (4) Download size doesn't match expected data structure. Critical for competitions where name suggests one format (e.g., "3D Surface Detection") but actual data is different (e.g., 2D images).

0
installs
▌
Kaggle Discussion Search · topprismdata
Use when looking for top-solution tricks on a Kaggle competition, when needing CV-LB gap discussions, when validating a strategy against community experience, or when seeking hyperparameter advice from competition winners. Triggers on phrases like "what did top scorers do", "CV-LB gap in this competition", "winning trick for X", or when starting any Kaggle competition.

0
installs
▌
LLM Competition Pipeline · topprismdata
Use when: (1) Entering an LLM-based Kaggle competition (open-source model required, prompt engineering, few-shot), (2) Need to choose between base models (Qwen, DeepSeek, LLaMA), (3) Designing a prompt strategy (zero-shot, few-shot, chain-of-thought), (4) Deciding whether fine-tuning is worth the GPU cost vs prompt engineering alone, (5) Setting up closed-book inference (no RAG, no internet). Covers the full pipeline: base model selection → prompt engineering → few-shot curation → evaluation → (optional) SFT/PEFT → submission. Inspired by NVIDIA's nemotron-customize pipeline structure, adapted for Kaggle LLM competition constraints.

0
installs
▌
Ptcg 2ply Bounded Search · topprismdata
Augment a BC Transformer agent with 2-ply bounded search using the cg library's SearchBegin/SearchStep API. Validated 2026-07-30 on Pokemon TCG AI Battle: search-enhanced agent submitted as v8 (LB pending). Use when: (1) you have a trained BC policy/value model, (2) the cg library exposes a search API (SearchBegin/SearchStep/SearchRelease), (3) you want lookahead beyond pure reaction-based action selection. Key risks: hidden-info sampling for unknown opponent state (use SNORLAX=1072 pad by convention), time budget per decision (N=5 samples × ~30 search_steps ≈ 1-3s), value-head quality unknown without calibration. Start with small N_SAMPLES and short rollout, tune up.

0
installs
▌
Autogluon Preset Strategy · topprismdata
AutoGluon preset selection strategy: when to use medium/good/high/best_quality, when EDA is unnecessary, and how to tune within a preset. Validated on s6e7 (AG high_quality 600s → OOF=0.8739, LB=0.87458, gap=0.0007 — perfect alignment). Use when: (1) Deciding which AutoGluon preset to start with, (2) Wondering whether to do manual EDA before fitting, (3) Wanting to know which hyperparameters to tune first, (4) Deciding time budget for tabular training. Differs from `autogluon-first` (which says "just use best_quality"): this skill explains the preset tradeoffs and where manual intervention pays off vs not.

0
installs
▌
Mem0 Dynamic Graph Memory · topprismdata
Use when extracting entities and relationships from text, when building dynamic knowledge graphs, when wanting to upgrade from static OKF to entity-aware memory, or when analyzing papers/discussions for concepts. Triggers when ingesting new papers, processing Kaggle writeups, or when existing skills can't capture connections between concepts.

0
installs
▌
Memory Hierarchy Management · topprismdata
Use when context window is getting crowded, when agent needs to recall past experiences, or when starting a long-running task that spans multiple sessions. Triggers when you have 43+ skills and can't fit them all in context, when you need to decide "what should be in working memory right now", or when persisting new learnings for future sessions.

0
installs
▌
Onnx Minimal Network Design · topprismdata
Use when: (1) Competing in ONNX-based competitions (NeuroGolf, model compression), (2) Need to design minimal neural networks as ONNX graphs, (3) Want to replace verbose agent-generated circuits with compact hand-designed nets. Key principles: single-node graphs have zero memory cost (output is free), Conv is a linear classifier (cannot separate nonlinear rules), Gather implements color permutations (10 params), use LogisticRegression to solve Conv weights for local 3×3 rules. Sparse initializers are rejected by ONNX strict checker for Conv nodes. Validated on NeuroGolf 2026 (7228 baseline, independent toolkit).

0
installs
▌
Tabular Polynomial Features Breakthrough · topprismdata
Polynomial features breakthrough for small tabular datasets. Use when: (1) Baseline model performance plateaus after hyperparameter tuning, (2) Dataset has <10K samples with <50 features, (3) Model optimization shows diminishing returns (<0.005 improvement), (4) Tree-based models (XGBoost/LightGBM) used but still underfitting. Covers degree-2 expansion capturing feature interactions that tree models miss, verified on ISEC 2026 competition (+0.017 LB improvement).

0
installs
▌
TS Day Specific Forecasting · topprismdata
Day-specific (direct) multi-step time series forecasting. Trains N separate models, one per prediction horizon day, with all features computed from the last known date. Eliminates stale lag feature problem entirely — predictions at correct magnitude without post-processing. CRITICAL: Do NOT apply geometric mean blending to day-specific model outputs — raw predictions are already at correct magnitude; blending DESTROYS accuracy (0.40 vs 3.2+). Use when: (1) Multi-step time series prediction (e.g., 16-day forecast), (2) Lag features become stale/constant during test, (3) Model underpredicts by 5-10x despite good CV, (4) Post-processing blends (geometric mean) are needed to fix magnitude in unified model. This is the 1st place approach from Favorita grocery sales competition. Applies to any multi-step time series forecasting with GBM/neural network models.

0
installs
▌
Agent Session Memory Protocol · topprismdata
Use when: (1) Starting a long-running experiment that may span multiple sessions/disconnects, (2) Resuming work after a break and needing to recover context, (3) Handing off to another agent or session, (4) You realize you've lost track of what was tried and why. Provides a structured 4-file session memory system that captures goal, timeline, files touched, and handoff instructions — lightweight (pure markdown, no infrastructure), works alongside AutoMem or standalone. Inspired by NVIDIA's nemo-rl-session-memory, adapted for Kaggle competition workflows.

0
installs
▌
Autogluon Timeseries Strategy · topprismdata
AutoGluon TimeSeriesPredictor: special API and presets for time series forecasting (different from TabularPredictor). Validated on Store Sales (N=3M, 33 families × 54 stores × 1684 days): AG 1.5 Chronos-2 + Chronos + onpromotion covariates → **LB RMSLE 0.39525** (best historical, vs AG 1.4 0.41852, vs manual 3.0+). Use when: (1) Working on time series competitions (forecasting), (2) Have multi-series data with optional covariates (promotions, holidays, prices), (3) Want AG 1.5 Chronos-2 zero-shot OR fine-tuned, (4) Need to bypass HF download errors. Differs from `autogluon-preset-strategy` (which covers TabularPredictor). Key breakthrough: using `model_path=LOCAL_PATH` to bypass `hf-mirror.com` download errors.

0
installs
▌
Sc Tir Mathematical Reasoning · topprismdata
Self-Consistency with Tool-Integrated Reasoning (SC-TIR) for mathematical problem solving. Use when: (1) Building mathematical reasoning systems, (2) Competing in math competitions (AIMO, AIME, MATH), (3) Need precise calculation beyond text generation, (4) LLM outputs contain calculation errors. Covers: SC-TIR algorithm (N×M candidates), Python code execution, majority voting, answer extraction from LaTeX \boxed{}, champion methodology from Numina AIMO Progress Prize winner.

0
installs
▌
Three Layer Wisdom Extraction · topprismdata
Use when: (1) User says "总结经验", "复盘", "extract methodology", "what did we learn", "抽象智慧", or asks to reflect on a completed project, (2) A non-obvious breakthrough, major failure, or counterintuitive discovery has just occurred, (3) User asks "why does this pattern keep appearing" or "what's the deeper principle here", (4) Before starting a new project when user wants to review past wisdom. Do NOT trigger for routine bug fixes — use claudeception instead.

0
installs
▌
Per Category Modeling Backfire · topprismdata
Per-category/per-family models can produce WORSE leaderboard scores than a single global model, even when per-category CV is better. Use when: (1) Considering training separate models per product category/family/store in tabular competitions, (2) Per-category CV improves but LB degrades, (3) Each category has <100K rows from a larger dataset. Covers data volume thresholds, hybrid fallback strategies, and when per-category modeling is appropriate vs counterproductive.

0
installs
▌
Ptcg Bc Large Model Submission · topprismdata
Submit a BC (behavioral cloning) Transformer agent for the Pokémon TCG AI Battle competition. Validated 2026-07-30: best.pth (15-epoch BC Large, d_model=256) → **public LB 600.0** (vs prior 165.6 — 3.6× improvement from using the best checkpoint instead of a stale submission). Use when: (1) training a BC agent on Kaggle Pokémon TCG episodes, (2) submitting a tar.gz agent for the competition, (3) scaling BC data to 80K+ games. Critical pitfall: your highest validation WR checkpoint is almost certainly NOT what you last saved to your submission directory — always verify the submission's model_bc.pth matches the best checkpoint before submitting. Stream-sampling (load 50K per part then drop) prevents OOM when scaling to 500K+ samples on a 64GB machine.

0
installs
▌
Rmsle Zero Threshold Asymmetry · topprismdata
Use when: (1) optimizing post-processing thresholds for RMSLE-evaluated competitions, (2) considering zeroing out small predictions, (3) implementing min-sales or adaptive thresholds for time series forecasting, (4) CV improves but LB degrades after changing post-processing, (5) comparing "smart" vs "simple" zeroing strategies. CRITICAL: Never assume that a "smarter" post-processing threshold is better for RMSLE without controlled experiment verification.

0
installs
▌
Ensemble Model Correlation Trap · topprismdata
Diagnose when ensemble/stacking will fail due to high model correlation. Use when: (1) Stacking or ensemble shows no improvement over single models, (2) Considering weighted average of multiple models, (3) Model predictions are nearly identical, (4) AUC/RMS gains plateau despite complex ensembling. Critical for tabular ML competitions where model diversity determines ensemble success.

0
installs
▌
Skill Governance And Automation · topprismdata
Use when: (1) Your skill repository has grown to 30+ skills and quality is inconsistent, (2) You need to audit skills for stale references, broken links, or missing evidence, (3) You want to automatically generate standardized summary cards for each skill, (4) You're onboarding a new agent and need a quick inventory of available skills. Provides: skill audit checklist (6 quality dimensions), automated skill card template, staleness detection protocol, and skill lifecycle management (create → validate → maintain → deprecate). Inspired by NVIDIA's skill-card-generator, adapted for community skill repositories.

0
installs
▌
Store Sales Darts Chronos Blend · topprismdata
Blend a foundation model (Chronos-2) with a tree model (darts LightGBMModel) for Kaggle Store Sales time-series forecasting. Validated 2026-07-26: Chronos-2 ensemble (LB 0.39387) + darts per-family LightGBM top-1 method (LB 0.39953) → geometric blend w=0.55 → **LB RMSLE 0.38444** (best, -0.012 vs single-model). Use when: (1) Forecasting competitions with multi-series data, (2) Single strong model has plateaued and you need a breakthrough, (3) You can train two INDEPENDENT algorithm-family models (neural + tree). Key insight: same-family blends are useless (Chronos v1+v2 correlation >0.99 → no gain), but cross-family blends give large gains even when both models score similarly. Differs from `autogluon-timeseries-strategy` (single Chronos-2 route) — this skill covers the multi-model BLEND that breaks the single-model ceiling.

0
installs
▌
Controlled Submission Experiment · topprismdata
Use when: (1) a new submission scores worse than baseline and the reason is unclear, (2) multiple changes were made simultaneously (new model + new post-processing + new features), (3) need to isolate which component caused a regression, (4) CV improves but LB degrades, (5) comparing "smart" vs "simple" approaches, (6) verifying that model improvements are real vs artifacts of post-processing changes. CRITICAL: Never change more than one variable between submissions without controlled comparison.

0
installs
▌
Kaggle Competition Type Strategy · topprismdata
Use when: (1) Entering a new Kaggle competition and unsure what type it is, (2) Need to choose between forking public kernels vs building custom, (3) Deciding how to allocate submission quota across the competition lifecycle, (4) Determining which evaluation/validation strategy to use. Covers 6 competition types (Standard Tabular, Code Competition, Simulation, Research, Prediction, Playground), each with distinct: submission mechanism, scoring system, public kernel value, GPU requirements, and optimal strategy. Includes a decision tree for first-48h actions and a quota allocation framework. Validated across 20+ competitions Jun-Sep 2026.

0
installs
▌
Kaggle Top Performer Replication · topprismdata
Systematically replicate and learn from top Kaggle performers. Use when: (1) Stuck at a plateau and need new ideas, (2) Want to understand how top performers achieved their scores, (3) Looking for proven techniques from successful submissions, (4) OOF keeps improving but LB stays same or gets worse. Covers: code analysis workflow, Frequency+Target encoding, Meta-learning with rank-transformed predictions, Temperature Scaling, and Multi-seed training.

0
installs
▌
Domain Knowledge Constraints Trap · topprismdata
Avoid applying domain knowledge constraints that hurt ML model performance. Use when: (1) Considering medical/physical/logical constraints on training data, (2) Adversarial validation AUC changes significantly after applying constraints (>0.10 shift), (3) CV score drops after adding "reasonable" domain rules, (4) Train/test distributions become misaligned after filtering.

0
installs
▌
Kaggle Competition Best Practices · topprismdata
Kaggle竞赛最佳实践和RAG知识库管理。使用时: (1) 开始任何Kaggle竞赛项目前建立规划, (2) 需要查询竞赛策略、获胜方案、特征工程技巧, (3) 需要为竞赛创建NotebookLM知识库。涵盖完整的竞赛工作流:从数据探索、特征工程、模型训练到提交策略。包含Stage1/Stage2 规则、获胜方案解析、ELO/Massey排名、评分函数优化等核心知识。

0
installs
▌
Kaggle Oof Lb Validation Protocol · topprismdata
Use when: (1) Your OOF score improved but LB didn't (or got worse), (2) You need to decide whether to submit based on OOF alone, (3) You're comparing models for ensemble inclusion and unsure which OOF signal to trust, (4) You see OOF/LB gap larger than 1% and don't know if it's overfitting or distribution shift. Covers: 5 sources of OOF/LB gap, the "OOF → LB confirmation protocol" (submit only after OOF + adversarial validation + format check), asymmetric gap patterns (OOF optimistic on small data, pessimistic on large), and the "3-strike rule" (if 3 consecutive OOF improvements don't improve LB, stop tuning and pivot). Validated across 20+ competitions including tabular (House Prices, SST, s6e7), simulation (PTCG), code (ROGII, Biohub, NeuroGolf), and time series.

0
installs
▌
Code Competition Artifact Pipeline · topprismdata
Use when: (1) Entering a Kaggle Code Competition that requires notebooks with no internet access, (2) Need to fork a public baseline that depends on external artifact datasets (pre-trained models, feature matrices, wheels), (3) Your fork fails with "module not found" or "file not found" errors, (4) Need to identify and attach the correct Kaggle datasets as notebook inputs. Key lessons: search kaggle datasets list for artifact bundles, check kernel-metadata.json dataset_sources field, the notebook's find_artifacts() function reveals expected paths. Validated on ROGII (7159→Pipeline A), Biohub (LB810 fork), NeuroGolf (7228 fork).

0
installs
▌
Cross Competition Feature Transfer · topprismdata
Use when starting a new ML task that structurally resembles a known competition (recommendation ↔ retail, demand ↔ supply, churn ↔ fraud, segmentation ↔ classification). Triggers when your task feels "not novel" — there is likely a top solution you can borrow features from. Especially valuable at the start of a project before inventing custom features.

0
installs
▌
Kaggle Cognitive Cost Optimization · topprismdata
Use when: (1) You have limited daily submission quota and 50+ skill files to explore, (2) Most Kaggle experiments cost 1 quota each but only the 1-2 BEST actually matter for your final score, (3) You waste hours doing redundant experiments because the EXISTING public kernels (rank 4-9 players) have already solved 80% of the problem, (4) You're tempted to re-submit "to see what happens". Key insight from 4+ months, 20+ competitions: **0.6 × BEST_PUBLIC ≥ 0.8 × OPTIMAL_OWN**, i.e. the best public kernel + small tuning > optimal own work in 90% of cases. Most Kaggle competition success comes from picking the right public kernel and spending quota wisely, not from complex custom work.

0
installs
▌
Kaggle Submission Format By Metric · topprismdata
Match the submission file format to the competition's evaluation metric BEFORE submitting. For ranking-based metrics (AUC, log_loss, MAP, NDCG, RMSLE) you MUST submit continuous probability/score values, not 0/1 class labels or rounded integers. Use when: (1) Preparing the final submission for any Kaggle competition, (2) AutoGluon's default `predict()` returns class labels for classification (must use `predict_proba()`), (3) Unsure whether to threshold predictions, (4) CV score is excellent but LB score is near-random. Real failure case: S6E2 Heart Disease rerun 2026-06-14 — OOF AUC 0.95554 dropped to Public LB 0.88403 with 0/1 submission, recovered to 0.95357 with probability submission.

0
installs
▌
Ladder Drift And Meta Aware Regression · topprismdata
Use when: (1) Your agent's LB score drops dramatically week-over-week without code changes, (2) You're tempted to add "meta-aware" bonuses (priority targets, matchup bonuses, deck-type detection) to a strong heuristic, (3) Testing agent variants on small samples (60-80 games) and getting noisy/conflicting results, (4) You see error rates spike when testing against meta decks. Do NOT trigger for standard tabular ML — this is about competition agents where (a) the meta shifts daily, and (b) "improvements" that look good in small tests often regress in production. Key lesson: same code, 200-point LB drift in 48h. Key anti-pattern: +600 target bonus on hypothetical Pokemon → -10% vs that deck.

0
installs
▌
Progressive Verification Debugging · topprismdata
Progressive verification methodology for debugging complex systems. Use when: (1) Facing mysterious crashes or errors with no clear cause, (2) System works in some environments but not others, (3) Multiple potential failure points exist, (4) Need to isolate whether issue is environment configuration vs code bug. Core principle: Start with simplest test, gradually increase complexity to pinpoint exact failure point.

0
installs
▌
Multi Level Aggregation Overfitting · topprismdata
Use when: (1) considering adding family-level or store-level aggregation features to day-specific models, (2) tempted to copy hierarchical features from 1st place solutions, (3) CV improves but LB degrades after adding aggregated features. WARN: Multi-level aggregation features that work in unified models may OVERFIT in day-specific (direct) forecasting frameworks.

0
installs
▌
Time Series Walk Forward Validation · topprismdata
Use when working on time series forecasting, sequence prediction, or any task with temporal data (weekly/daily orders, retail SKU recommendation, demand forecasting, user churn prediction). Triggers when you need to set up cross-validation, evaluate a temporal model, or audit temporal leakage. Always use instead of K-fold on time-indexed data.

0
installs
▌
Unified Vs Day Specific Forecasting · topprismdata
Use when: (1) deciding between unified vs day-specific models for multi-step forecasting, (2) considering blending day-specific with unified predictions, (3) CV shows day-specific worse than unified but unsure if LB will agree, (4) building multi-step time series models with >7 day horizons. KEY FINDING: Unified model with target_day_offset feature can significantly outperform day-specific models, especially when features are rich enough to encode temporal patterns.

0
installs
▌
Competition Orchestration Multimodel · topprismdata
Use when: (1) You need to rapidly iterate across MANY competing approaches in a single Kaggle competition (NeuroGolf-style: 7+ different public kernels forked in <2 hours), (2) You must decide which public dataset/kernel to spend your limited submission quota on, (3) Your highest-scoring submission might be a public dataset's submission, not your own engineered one, (4) You see sub-1.0 "0.0 placeholder" submissions that look like errors but are actually correct. Key lessons: from 7228→7269.68 (+41) in 4 hours via public kernel forking; from 0.0→62.64 in 12h via TensorLiu's BUDGET-FILLING strategy; submission batching is essential to avoid "pending" serialization. Validated against NeuroGolf 2026, AI Agent Security 2026-07, Biohub Cell Tracking (0.0→pending), PTCG AI Battle (final-2 strategy).

0
installs
▌
TS Forecasting Stale Lag Methodology · topprismdata
Complete methodology for solving multi-step time series forecasting competitions where lag features cause systematic underprediction at test time. Documents the full journey from LB 1.859 → 0.399, including every failed approach and the key breakthrough. Use when: (1) Starting a new time series forecasting competition, (2) Model underpredicts at test time due to stale lag features, (3) CV-LB gap of 3-10x persists, (4) Need a systematic approach to diagnose and fix prediction magnitude issues. This is the MASTER methodology combining all ts-* skills.

0
installs
▌
Adversarial Validation Kaggle · topprismdata
Correct implementation of adversarial validation for Kaggle competitions. Use when: (1) Selecting training samples that match test distribution, (2) Reducing synthetic data artifacts by filtering, (3) Preparing data subsets for better generalization. Covers train vs test classification, sample selection method, and common pitfalls.

0
installs
▌
Knowledge Crystallization Feedback Loop · topprismdata
Use when: (1) You've completed a competition or experiment and want to extract reusable knowledge, (2) Your agent keeps repeating the same mistakes across competitions because lessons weren't crystallized, (3) You have 50+ memory files but can't find relevant knowledge when starting a new task, (4) You need to decide what to remember vs what to forget. Covers the complete crystallization cycle: Experiment → Identify Outcome → Extract Pattern → Classify (feedback/ learned/project/reference) → Store with triggers → Activate on match. Includes the "3-layer architecture" (L1 core, L2 domain skills, L3 cross-domain principles) and the "forgetting protocol" for stale knowledge. Validated across 4+ months, 20+ competitions, 109+ crystallized memories.

0
installs
▌
Feature Engineering Saturation Detection · topprismdata
Use when you have run 3+ feature engineering experiments with no improvement, or when distance to theoretical upper bound is <15pp. Triggers when you suspect "I should stop optimizing features and try something else" — the most common failure mode is wasting weeks past the saturation point.

0
installs
▌
Spatiotemporal Graph Feature Engineering · topprismdata
Feature engineering patterns for spatiotemporal graph prediction tasks. Use when: (1) Working with time series on graph structures, (2) Predicting node-level outcomes over time, (3) Network dynamics with spatial dependencies, (4) Kaggle competitions involving traffic, flood, power grid, or social networks. Covers: neighbor aggregation, temporal windows, cross-domain features.

0
installs
▌
Trueskill Simulation Competition Strategy · topprismdata
Use when: (1) Competing in Kaggle simulation competitions (PTCG, Orbit Wars, etc.) where agents are rated by TrueSkill Bayesian scoring, (2) Your LB score swings ±200 points day-to-day with no code changes, (3) You're tempted to re-submit frequently to "test" improvements, (4) You see identical agents scoring 150-400 points differently. Key lessons: re-submitting resets μ to 600 (destroys convergence), only latest 2 submissions count for final, the optimal endgame strategy is "submit your best agent's duplicate copies near deadline" to high-roll the Bayesian lottery. Validated across 8+ PTCG submissions + forum consensus from rank 4-9 players.

0
installs
▌
Learned Value Beats Heuristic Augmentation · topprismdata
Use when: (1) You have a strong hand-crafted baseline (heuristic, rule-based agent, domain-expert system) and want to improve it with ML, (2) Naive augmentation (search, behavioral cloning, type-classification) is making things WORSE, not better, (3) You're deciding between "more training data" vs "a learned value function" for a game/RL/problem, (4) A competition baseline resists every improvement attempt. Do NOT trigger for tabular/standard ML where the baseline is itself a learned model — this is about beating a STRONG HEURISTIC with learned components.

0
installs

64 published skills