Agent Skills — Kaggle Grandmaster (Machine Learning Competitivo)
OlixIgnacious/agent-skills
12 skills Kaggle GM: EDA → features → ensemble → submissão
Ver no GitHubBiblioteca de skills e agents para Claude Code com 12 comandos especializados em competições Kaggle. Cobre todo o fluxo de trabalho: validação adversarial, EDA, baselines diversificados, otimização com Optuna, engenharia de features, hill climbing, stacking, pseudo-labeling e ensemble com seed. Instalável em macOS, Linux e Windows via PowerShell.
Instalação
npx convez-ai OlixIgnacious/agent-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Otimizar modelos em competições Kaggle com orquestração automática
- Implementar validação cruzada segura e detecção de data drift
- Engenharia de features acelerada com GPU e rastreamento de delta de CV
- Busca bayesiana de hiperparâmetros e otimização de pesos de ensemble
Veredito

Ferramenta muito especializada para competidores Kaggle serious. Implementação sólida com agents bem estruturados, mas requer expertise em ML competitivo para máximo proveito. Excelente para acelerar fluxo de validação e ensemble.
README
A tradução para o português deste README ainda está na fila.
Agent Skills
A multi-domain Claude Code skill library — agents and skills organized by use case. Install once, use what's relevant.
macOS / Linux
bash <(curl -fsSL https://raw.githubusercontent.com/OlixIgnacious/agent-skills/main/install.sh)
Windows (PowerShell)
irm https://raw.githubusercontent.com/OlixIgnacious/agent-skills/main/install.ps1 | iex
The installer asks which domain (Kaggle / SDLC / Research / All) and which tool (Claude Code / Antigravity / Copilot / All), then downloads only the files you need.
By default, Claude Code skills are installed into ~/.claude/skills/ so they work in every project. Pass --local to install into the current project's .claude/skills/ instead:
bash <(curl -fsSL https://raw.githubusercontent.com/OlixIgnacious/agent-skills/main/install.sh) --local
How It Works
Agents are role-based AI workers Claude delegates to automatically based on your task. They have a specific persona, judgment filters, and tool access.
Skills are slash commands you invoke explicitly for step-by-step procedures with code templates.
/kaggle-grandmaster ← you invoke a skill
"help me engineer features" ← Claude delegates to an agent
Domains
Kaggle — Competitive Machine Learning
3 agents · 12 skills · Full GM workflow from EDA to final submission
Agents
| Agent | Role |
|---|---|
kaggle-grandmaster | Competition orchestrator — decides phase, invokes skills, delegates to sub-agents |
kaggle-feature-engineer | Feature engineering specialist — CV-delta tracked batches, GPU-accelerated |
kaggle-ensemble-builder | Ensemble specialist — hill climbing, stacking, Optuna weight search, final blend |
Skills
| Command | Phase | What it does |
|---|---|---|
/kaggle-grandmaster | Always | Persona, shake-up avoidance, problem reformulation heuristics |
/kaggle-adversarial-validation | Day 1 | Detect train/test distribution shift before locking folds |
/kaggle-validation | Day 1 | Lock fold strategy — KFold / TimeSeriesSplit / GroupKFold / Multilabel |
/kaggle-eda | Day 1–2 | Distribution analysis, leakage scan, temporal pattern detection |
/kaggle-baselines | Day 2–3 | Ridge + XGBoost + LightGBM + CatBoost with GPU backends |
/kaggle-target-transform | Day 2–3 | Log/sqrt transforms, winsorization, isotonic calibration |
/kaggle-optuna | Day 3+ | Bayesian hyperparameter search (TPE) + ensemble weight optimization |
/kaggle-feature-engineering | Day 3→N-5 | Groupby aggs, interactions, CV-safe target encoding, lag/rolling |
/kaggle-hill-climbing | Final week | Greedy ensemble selection + scipy/Optuna weight optimization |
/kaggle-stacking | Final week | Stage 1 OOF → Stage 2 meta-learner + residual stacking |
/kaggle-pseudo-labeling | Final 3 days | Fold-aware soft labels on unlabeled test data |
/kaggle-extra-training | Final 24h | Seed ensemble + full-data retrain + submission checklist |
→ Domain guide · Orchestration
SDLC Workbench — Software Development Lifecycle
17 agents · 4 skills · Requirements → Architecture → Implementation → Review
Skills
| Command | Phase | What it does |
|---|---|---|
/sdlc-biz-to-tech | Phase 1 | Business requirement → TRD grounded in the codebase |
/sdlc-architectural-review | Phase 2 | Stress-test a TRD — scalability, security, data, ops |
/sdlc-feature-dev | Phase 3 | Implement from approved TRD — TDD, full test coverage |
/sdlc-code-review | Phase 4 | Multi-dimensional PR review → APPROVE / REQUEST CHANGES / BLOCK |
A drop-in workflow template for any codebase. Works across Claude Code, Antigravity, and GitHub Copilot.
| File | Tool |
|---|---|
CLAUDE.md | Claude Code |
AGENTS.md | Universal — Antigravity + Cursor + Claude Code |
GEMINI.md | Antigravity (agy) |
.github/copilot-instructions.md | GitHub Copilot |
Research Paper Writing
2 agents · 1 skill · Venue selection → Literature review → Writing → Submission
Agentic research assistant that searches the web for top papers, fetches live journal/publisher formatting guidelines, and guides writing from abstract to references. Adapts to any venue.
Agents
| Agent | Role |
|---|---|
literature-reviewer | Deep literature search — arXiv, Semantic Scholar, venue proceedings. Annotated bibliography + gap analysis |
venue-advisor | Identifies and ranks target journals/conferences. Fetches live guidelines, impact factors, acceptance rates |
Skills
| Command | What it does |
|---|---|
/research-paper | End-to-end paper writing — venue selection, literature search, section-by-section writing, publisher formatting, submission checklist |
Supported publishers: IEEE, ACM, Nature, Springer (LNCS), Elsevier, NeurIPS, ICML, ICLR, ACL, CVPR, arXiv and more — guidelines fetched live.
Finance Research (coming soon)
1 agent · skills planned · Quant research, factor analysis, backtesting, portfolio construction
Agents
| Agent | Role |
|---|---|
finance-researcher | Quantitative research — factor IC analysis, walk-forward backtesting, risk |
Meta
1 skill · Build new skills and agents for this library
| Command | What it does |
|---|---|
/skill-creator | Guided creation of a new skill or agent — gathers requirements, writes correctly structured files, places them in the right directories, updates the README |
Repository Structure
agents/ ← role-based AI workers (auto-delegated)
kaggle/
kaggle-grandmaster.md
kaggle-feature-engineer.md
kaggle-ensemble-builder.md
sdlc/
biz-to-tech-orchestrator.md
architectural-review-orchestrator.md
feature-dev-orchestrator.md
code-review-orchestrator.md
requirements-analyst.md
code-archaeologist.md
api-designer.md
test-engineer.md
technical-writer.md
software-engineer.md
database-internals.md
devops-sre.md
cybersecurity.md
linux-debugging.md
system-design.md
competitive-programming.md
ml-research.md
research/
literature-reviewer.md
venue-advisor.md
finance/
finance-researcher.md
skills/ ← slash-command skills, domain-prefixed
kaggle-grandmaster/
kaggle-adversarial-validation/
kaggle-validation/
kaggle-eda/
kaggle-baselines/
kaggle-target-transform/
kaggle-optuna/
kaggle-feature-engineering/
kaggle-hill-climbing/
kaggle-stacking/
kaggle-pseudo-labeling/
kaggle-extra-training/
sdlc-biz-to-tech/
sdlc-architectural-review/
sdlc-feature-dev/
sdlc-code-review/
research-paper/
skill-creator/
domains/ ← per-domain workflow guides and orchestration
kaggle/
sdlc/
research/
finance/
.claude-plugin/ ← plugin manifest
plugin.json
Contributing a Domain
agents/<domain>/<domain>-<role>.md— agent withname,description,model,toolsfrontmatterskills/<domain>-<name>/SKILL.md— skill withdescription,disable-model-invocation,allowed-toolsfrontmatter (omitdisable-model-invocationfor persona/orchestrator skills)domains/<domain>/README.md— use case overview, file list, and skill chaindomains/<domain>/ORCHESTRATION.md— phase-by-phase workflow (optional — see Kaggle for an example)
See the Kaggle domain as the reference implementation for skills; the SDLC domain for agents.