Claude ML Intern Skill automates research workflows from planning through training
A new GitHub skill wires Claude into an autonomous research loop that plans implementations, debugs code, and trains models without human hand-holding — already used to train a DeepSeek-4-style architecture and port Flash Attention to Volta GPUs.
Claude ML Intern Skill is an autonomous research agent framework that handles the full loop of machine learning experimentation: planning, resource discovery, debugging, and training. The skill addresses a recurring pain point with Claude Code and similar AI coding assistants — the tendency to roll back changes, leave implementations half-finished, or ask for manual intervention mid-task.
The framework plugs into Claude's API and runs the agent through a structured workflow. It searches for relevant papers and codebases, writes implementation plans, executes training runs, and iterates on bugs autonomously. The GitHub repository includes setup instructions and example workflows for common ML tasks.
Deployed examples
The creator has used the skill to train a 100M-parameter model following the DeepSeek-4 architecture on the Tiny Stories dataset. The trained checkpoint is live on HuggingFace Spaces as a working demo. The agent also ported Flash Attention to Volta-generation GPUs — a non-trivial CUDA kernel adaptation that typically requires manual debugging across multiple build cycles. Both projects ran with minimal human supervision once the agent was pointed at the task.
The skill is open-source and actively maintained. Practitioners looking to offload repetitive research loops — hyperparameter sweeps, ablation studies, architecture ports — can clone the repository and wire it into their own Claude workflows. The code is Python-based and assumes access to the Claude API.
