agent-skills

Plugin agent-skills in OlixIgnacious/agent-skills. Commands read from the repo's manifest — 1 month ago.

Skills (18)

• kaggle-adversarial-validation

  Detect train/test distribution shift by training a classifier to distinguish train from test rows. AUC near 0.5 means distributions match and CV is trustworthy. High AUC signals your CV will mislead you. Run before locking any fold strategy.

• kaggle-baselines

  Build diverse model baselines simultaneously for Kaggle competitions — linear models, GBDTs (XGBoost, LightGBM, CatBoost), and neural nets. GPU-accelerated. Establishes the CV score floor before any feature engineering.

• kaggle-eda

  Smarter EDA for Kaggle competitions. Goes beyond basic statistics to detect train/test distribution shift, temporal patterns, target leakage signals, and data quality issues that directly affect modeling decisions.

• kaggle-extra-training

  Final competition preparation — seed ensembling for variance reduction, full-data retraining after hyperparameter lock, and pre-submission checklist. This is the last skill to run before final submission.

• kaggle-feature-engineering

  Feature engineering at scale for Kaggle competitions. Covers groupby aggregations, interaction features, target encoding (CV-safe), lag/rolling features for time series, and GPU-accelerated feature generation with cuDF. Run after baselines to measure delta impact of each feature batch.

• kaggle-grandmaster

  Activates Kaggle Grandmaster thinking mode. Load this first at the start of any competition session to set judgment filters, decision heuristics, and communication style for all subsequent Kaggle skills.

• kaggle-hill-climbing

  Build an ensemble by greedy forward selection — start with the best single model and add others only when the metric improves. Uses CuPy for vectorized metric computation and scipy for weight optimization.

• kaggle-optuna

  Hyperparameter optimization with Optuna's TPE sampler for LightGBM, XGBoost, CatBoost, and neural nets. Also covers Optuna-guided ensemble weight optimization as a superior alternative to scipy hill climbing for large model pools.

• kaggle-pseudo-labeling

  Add high-confidence test predictions as soft labels to the training set, then retrain. Iterative multi-round approach for competitions with unlabeled test data. Most effective in semi-supervised settings with large test sets.

• kaggle-stacking

  Multi-level stacking for Kaggle — Stage 1 generates out-of-fold predictions as meta-features, Stage 2 trains a meta-learner on them. Produces more powerful ensembles than weighted averaging alone.

• kaggle-target-transform

  Target and prediction transformations for Kaggle — log/sqrt target transforms for skewed regression targets, winsorization for outlier robustness, isotonic regression for probability calibration, and beta-sharpening for confidence-boosting predictions. Apply before training and reverse after.

• kaggle-validation

  Design and lock the validation strategy for a Kaggle competition. Covers fold type selection (KFold, StratifiedKFold, TimeSeriesSplit, GroupKFold), CV-LB correlation verification, and out-of-fold prediction framework. Run this before any feature engineering or modeling.

• research-paper

  End-to-end research paper writing skill. Searches the web for top papers in your area, fetches journal/publisher formatting guidelines, identifies related work, and guides writing from abstract to references. Adapts to IEEE, ACM, Nature, Springer, Elsevier, NeurIPS, ICML, and more.

• sdlc-architectural-review

  Stress-test a TRD or design document before any code is written. Spawns system-design, cybersecurity, database-internals, and devops-sre in parallel for a multi-dimensional review. Produces an APPROVED or REVISION REQUIRED verdict with specific findings.

• sdlc-biz-to-tech

  Translate a business requirement, PRD, stakeholder ask, or vague goal into a grounded Technical Requirements Document (TRD). Delegates to biz-to-tech-orchestrator, which spawns requirements-analyst and code-archaeologist in parallel before drafting the TRD.

• sdlc-code-review

  Multi-dimensional PR review before merging — correctness, security, test coverage, and database safety. Spawns software-engineer, cybersecurity, test-engineer, and database-internals in parallel. Produces an APPROVE, REQUEST CHANGES, or BLOCK decision with findings ranked by severity.

• sdlc-feature-dev

  Implement an approved TRD — architecture, code, and full test coverage. Requires an approved TRD as input. Spawns api-designer, test-engineer, software-engineer, devops-sre, and technical-writer in sequence. Never starts coding without an approved design.

• skill-creator

  Creates new skills and agents for any tool in this library — Claude Code (SKILL.md + agents), Google Antigravity (AGENTS.md + GEMINI.md), and GitHub Copilot (.github/copilot-instructions.md). Gathers requirements, inspects existing examples, writes correctly structured files, places them in the righ

Similar plugins