Autoresearch (generalized)

Karpathy's loop in one breath: pick a goal, define a mechanical fitness function, constrain a mutable surface, iterate keep-or-revert until budget or plateau. This skill ports it to any domain.

The hard-won lesson: most candidate workflows are not autoresearch-shaped. Running the loop on a fat-tailed, reflexive, or slow-feedback problem converges to a local optimum of the wrong thing - very efficiently. This skill's main job is to catch that before the loop starts. Saying "this isn't autoresearch-shaped, here's why, here's what to do instead" is the correct output when a candidate fails triage.

When this skill applies

Invoke when the user:

• wants to scaffold or run an iterative improvement loop on any workflow
• mentions Karpathy, autoresearch, keep-or-revert, overnight optimization, Tobi's Liquid, pi-autoresearch, "the Karpathy loop"
• asks "can I apply autoresearch to X"
• has a measurable goal and wants autonomous iteration
• says "overnight agent," "experiment loop," "fitness function," "val_bpb"