Bootstrap autonomous ML research on any project. Explores the codebase, classifies directories into "prepare.py scope" (fixed infrastructure — data, metrics, callbacks) vs "train.py scope" (modifiable — models, training, losses, optimizers), then generates a program.md research guide adapted from karpathy/autoresearch. Use when (1) user wants to set up autonomous research on an ML project, (2) user mentions "autoresearch" or "program.md" in context of their project, (3) user wants to identify which code to modify vs keep fixed for experimentation. Works on any ML project regardless of framework.
2026-07-16