Can it help me avoid repeating failed experiments?

Yes, the skill is designed to scan your troubleshooting guides and past logs to identify known failure patterns and ensure your next plan avoids those specific issues.

How do I trigger the advisory behavior?

You can trigger the skill by starting your message with the tag or by asking natural language planning questions like 'What should we try next?'

What domains does the Domain Advisor support?

It currently specializes in three primary domains: ML Research (hyperparameters and architecture), Unsloth (LLM fine-tuning and quantization), and CUDA (kernel performance and memory optimization).

Where does the skill look for project history?

It reads from the .codex/skills/registry.json file to locate your specific paths for experiment logs, benchmark reports, and troubleshooting guides.

Domain Advisor

Name: Domain Advisor
Author: Erland366

byErland366

Data Science & ML

Plans data science and hardware optimization experiments by analyzing historical project logs and domain-specific configurations.

About

The Domain Advisor skill transforms Claude into an intelligent project strategist that leverages your repository's history to suggest the most logical next steps for your development. By reading from a centralized registry and analyzing past experiment logs, it provides tailored advice for machine learning research, Unsloth LLM fine-tuning, and CUDA kernel optimization. Whether you are optimizing hyperparameters or debugging memory usage, this skill generates structured experimental plans that respect your hardware constraints, reuse existing defaults, and avoid repeating past failures.

Key Features

Context-aware experiment planning based on historical logs and performance reports
0 GitHub stars
Automated analysis of troubleshooting guides to avoid known failure patterns
Specialized domain support for ML research, LLM fine-tuning (Unsloth), and CUDA optimization
Structured output with detailed experiment tables including objectives and parameters
Integration with project registries for consistent path and configuration management

Use Cases

Planning kernel optimization strategies and tiling configurations for CUDA development
Determining the next set of hyperparameter sweeps during model training
Optimizing memory usage and quantization settings for LLM fine-tuning with Unsloth

About

Key Features

Context-aware experiment planning based on historical logs and performance reports
0 GitHub stars
Automated analysis of troubleshooting guides to avoid known failure patterns
Specialized domain support for ML research, LLM fine-tuning (Unsloth), and CUDA optimization
Structured output with detailed experiment tables including objectives and parameters
Integration with project registries for consistent path and configuration management

Use Cases

Planning kernel optimization strategies and tiling configurations for CUDA development
Determining the next set of hyperparameter sweeps during model training
Optimizing memory usage and quantization settings for LLM fine-tuning with Unsloth