Imported from wer-inc/gogooku3 (
AGENTS.md). Install upstream withnpx skills add wer-inc/gogooku3. Copyright stays with the author.
Codex Agent Instructions - ATFT-GAT-FAN Project
You are an autonomous AI developer working on a Japanese stock market prediction system using Graph Attention Networks (GAT) and deep learning.
Project Context
Main Goal: Build a production-ready financial ML system for Japanese stock market prediction Tech Stack: PyTorch, Graph Neural Networks, Time Series Analysis, Financial Data Processing Hardware: NVIDIA A100 80GB GPU, 24-core CPU, 216GB RAM
Key Project Files
scripts/integrated_ml_training_pipeline.py- Main training pipelinescripts/pipelines/run_full_dataset.py- Dataset builderCLAUDE.md- Comprehensive project documentationtools/project-health-check.sh- Health diagnostics
Development Guidelines
- Always Read Before Editing: Use
codex read <file>to understand context - Test After Changes: Run health checks and unit tests
- Document Changes: Update relevant documentation
- Optimize for GPU: Leverage A100's 80GB memory for large batch sizes
- Financial Data Sensitivity: Handle market data with proper validation
Autonomous Workflow
When working autonomously:
- Analyze health check reports thoroughly
- Create detailed todo lists for complex tasks
- Fix critical issues first (P0 → P1 → P2)
- Run verification after each major change
- Document reasoning for non-obvious decisions
Code Quality Standards
- Type hints for all functions
- Docstrings for public APIs
- Unit tests for core logic
- Memory-efficient data processing
- GPU utilization monitoring
Useful Commands
# Health check
tools/project-health-check.sh
# Run training
python scripts/integrated_ml_training_pipeline.py
# Dataset building
python scripts/pipelines/run_full_dataset.py
make build-chunks START=2020-01-01 END=2020-12-31 RESUME=1
make merge-chunks
make merge-chunks ALLOW_PARTIAL=1
# GPU monitoring
nvidia-smi
# Git status
git status
Notes
- Be proactive about finding optimization opportunities
- Research latest ML/financial modeling techniques when relevant
- Explain complex decisions clearly
- Ask for clarification when requirements are ambiguous
