菜单
Create a new experiment folder by copying reference/ as a template. Auto-generates SESSION_NOTES.md and run.sh. Claude experiments use expA00, A01, ...; human experiments use exp200, 201, ... in sequential numbering.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
コンペ開始時の7項目チェックリストを対話的に埋め、survey/competition/overview.md として保存する。新しいコンペに着手するときに最初に使う。
提出物の事前検証を実行する。Kaggle CSV / Kaggle Code Competition / 予測ファイル zip / Docker コンテナ / Simulation エージェントの5形式に対応。submission-validator agent を呼び出して機械的にチェックする。提出直前に必ず使う。
Walks through the 7-item competition onboarding checklist interactively and saves it to survey/competition/overview.md. Use first when starting a new competition.
Run pre-submission validation. Handles 5 formats — Kaggle CSV / Kaggle Code Competition / prediction-file zip / Docker container / Simulation agent. Delegates to the submission-validator agent for mechanical checks. Use right before any submission.
今日の daily_reports/YYYYMMDD.md を、前日のレポートと未完了TODO・直近の実験結果を引き継いで作成する。セッション開始時に使う。
新しい実験フォルダを reference/ から雛形コピーして作成する。SESSION_NOTES.md と run.sh も自動生成。Claude用は expA00, A01, ...、人間用は exp200, 201, ... の連番で命名する。
| name | exp-new |
| description | Create a new experiment folder by copying reference/ as a template. Auto-generates SESSION_NOTES.md and run.sh. Claude experiments use expA00, A01, ...; human experiments use exp200, 201, ... in sequential numbering. |
| argument-hint | <experiment_name> [--human for human-style exp200_xxx] |
Parse arguments:
$ARGUMENTS (required)--human flagDetermine folder name:
workspace/expA00_<name>, advance to A01, A02, ..., B00, ... based on existing foldersworkspace/exp200_<name>, advance to 201, 202, ...ls workspace/ first to determine the next numberCopy template:
cp -r reference/ workspace/<new-folder>/
Generate SESSION_NOTES.md (don't overwrite if it exists):
# SESSION_NOTES — <folder name>
## Session Info
- Date: YYYY-MM-DD
- Working folder: workspace/<folder name>
- Goal: <inferred from name or asked from user>
## Base
- Base: reference/ (PyTorch Lightning + timm + smp)
- Parent exp (if any): ...
## Approaches Tried and Results
<!-- Append after each run -->
## File Structure
- train.py
- src/
- config.yaml
- run.sh
## Key Insights
<!-- Insights from training -->
## Next Steps
- [ ] ...
## Performance History
| Trial | Change | CV | LB | Notes |
|-------|--------|-----|-----|-------|
## Command History
<!-- Major commands run -->
#!/bin/bash
# Train + inference in one script
# IMPORTANT: cd "$(dirname "$0")" must run before python.
# Without it, results/ may be created at the repo root or elsewhere.
# Usage: bash run.sh data.fold=0 trainer.max_epochs=30
# Arguments override config.yaml in OmegaConf dotlist format (there is no --config flag)
set -e
cd "$(dirname "$0")"
# Train (train.py lives directly in the experiment folder; pass arguments through)
python train.py "$@"
# Inference (as needed)
# python predict.py ckpt_path=results/<experiment_name>/fold0/last.ckpt
Update experiment.name in config.yaml to the new folder name
Make the results/ output path relative to the experiment folder:
train.py's output path is not cwd-dependentoutput_dir = Path(__file__).resolve().parent / 'results' / experiment_name / f'fold{fold}' (one parent because train.py lives directly in the experiment folder)workspace/<new-folder>/results/... before reporting to the userReport:
bash workspace/<new-folder>/run.sh).gitignore)