The Agent Skills Directory

[SAFE]: The skill implements a structured workflow for research automation with several safety mechanisms.
[SAFE]: Requires all work to be performed on isolated Git branches (exp/*) to prevent data loss or main branch corruption.
[SAFE]: Mandates logging of all experimental results to a local file (experiments.jsonl) for transparency and auditability.
[SAFE]: Includes explicit safety instructions to never delete data, restrict file modifications to the project scope, and adhere to cost limits.
[SAFE]: Establishes a 'human-in-the-loop' requirement where the user must approve prompt changes and direction shifts.
[SAFE]: Provides a clear stopping condition if the evaluation metric degrades consistently.

karpathy-autoresearch