buzziru/agentic-kaggle-tabular-template
Agentic Kaggle tabular experimentation template — OOF CV, stacking, leakage guards & post-mortem-driven workflow. Claude Code native.
ARCHETYPE
Orchestrator
A bench of specialised subagents. The main agent mostly delegates.
Copy this rig
# review before running: this installs third-party code
$ npx degit buzziru/agentic-kaggle-tabular-template/.claude ./rig-agentic-kaggle-tabular-template # inspect, then merge into .claude/ MCP servers are added to Claude Code at local scope; env vars are shown as YOUR_… placeholders — we never store values. Files are fetched with degit into a separate folder so you can review before merging.
$ curl -fsSL --create-dirs -o .claude/agents/code-reviewer.md https://raw.githubusercontent.com/buzziru/agentic-kaggle-tabular-template/main/.claude/agents/code-reviewer.md $ curl -fsSL --create-dirs -o .claude/agents/eda-explorer.md https://raw.githubusercontent.com/buzziru/agentic-kaggle-tabular-template/main/.claude/agents/eda-explorer.md $ curl -fsSL --create-dirs -o .claude/agents/exp-runner.md https://raw.githubusercontent.com/buzziru/agentic-kaggle-tabular-template/main/.claude/agents/exp-runner.md $ curl -fsSL --create-dirs -o .claude/agents/feature-smith.md https://raw.githubusercontent.com/buzziru/agentic-kaggle-tabular-template/main/.claude/agents/feature-smith.md $ curl -fsSL --create-dirs -o .claude/agents/premise-auditor.md https://raw.githubusercontent.com/buzziru/agentic-kaggle-tabular-template/main/.claude/agents/premise-auditor.md $ curl -fsSL --create-dirs -o .claude/agents/result-reviewer.md https://raw.githubusercontent.com/buzziru/agentic-kaggle-tabular-template/main/.claude/agents/result-reviewer.md
Merge into .claude/settings.json (project) or ~/.claude/settings.json (user). Hook commands reference scripts in the source repo — copy those too.
{
"hooks": {
"PreToolUse": [
{
"matcher": "Edit|MultiEdit|Write|NotebookEdit",
"hooks": [
{
"type": "command",
"command": "python scripts/hooks/guard_frozen.py"
}
]
},
{
"matcher": "Bash",
"hooks": [
{
"type": "command",
"command": "python scripts/hooks/guard_bash.py"
},
{
"type": "command",
"command": "python scripts/hooks/remote_run_reminder.py"
}
]
}
]
}
} Subagents (6)
| code-reviewer model: opus | 작성·수정된 실험/공유 코드를 고정 체크리스트로 검증한다. 구현 직후, exp-runner 실행(풀 학습/push) 전에 사용한다. self-review 관대함을 차단하려고 작성자와 분리한다. |
| eda-explorer model: sonnet | Read-only EDA agent. Use when you need to explore data (distributions, target relationships, group/leakage checks, train/test drift) and want only a concise numeric summary back, not verbose DataFrame |
| exp-runner model: sonnet | 헤드리스 실험 실행 에이전트. `src/` 코드를 원격(기본 Kaggle GPU)에 push 하고 노트북/잡을 서버에서 실행·모니터·산출물 회수한다. 로컬에 GPU 가 없을 때 대형/신경망 모델을 돌릴 때 사용. **블로킹 금지** — push·RUNNING 확인 후 빠르게 리턴하고, 완료 회수는 별도 호출. |
| feature-smith model: opus | Feature engineering agent. Use when implementing or revising features in src/features.py. It implements features, runs a leakage check, and measures OOF impact against the baseline. Use one at a time |
| premise-auditor model: opus | 판정 5건마다 현재 방향의 전제를 공격하는 blind 감사를 수행한다. 누적 결과 숫자만 입력으로 받고, 설계 rationale·서사는 읽지 않는다. kill/continue 판정은 사람의 몫. |
| result-reviewer model: opus | 실행 완료된 실험의 로그를 사전 등록된 expectation 과 대조해 판정 하나를 내린다. 풀 실행 완료 직후 사용한다. 다음 실험을 제안하지 않는다(제2의 설계자 방지). |
Hooks (5)
| event | matcher | runs |
|---|---|---|
| PreToolUse | Edit|MultiEdit|Write|NotebookEdit | python scripts/hooks/guard_frozen.py |
| PreToolUse | Bash | python scripts/hooks/guard_bash.py |
| PreToolUse | Bash | python scripts/hooks/remote_run_reminder.py |
| PostToolUse | NotebookEdit | python scripts/hooks/lint_notebook.py |
| Stop | * | python scripts/hooks/stop_reminder.py |
Similar rigs
AizenvoltPrime/claude-setup
—
Pragmatist 1.2k tok ·
IvanKuzyshyn/dotfiles
Dotfiles optimized for speed, coffee, and questionable life choices. 😜
Pragmatist 3.0k tok ·
HGInsights/claude-code-setup
Production-tested Claude Code harness configuration: hooks, skills, subagents, and automated review loops
Pragmatist 1.9k tok ·
stavrespasov/claude-code-os-lite
A free, installable starter setup for Claude Code: 10 workflow commands, 2 agents, and the hook that ends --no-verify commits. MIT.
Pragmatist 101 tok ·