The Agent Skills Directory

[SAFE]: The skill describes the use of PufferLib, a legitimate and well-known reinforcement learning library. All code examples use standard PyTorch and PufferLib patterns.
[COMMAND_EXECUTION]: Provides standard training commands using torchrun and puffer train. These are common developer workflows for distributed training and do not involve unauthorized privilege escalation or dangerous execution.
[EXTERNAL_DOWNLOADS]: The skill recommends installing PufferLib via uv pip install pufferlib, which is a standard package management operation for a known library.
[PROMPT_INJECTION]: No evidence of instructions attempting to bypass safety filters or override agent behavior was found.
[DATA_EXFILTRATION]: There are no patterns suggesting unauthorized data access or network transmission of sensitive information. Mention of logging services like Weights & Biases or Neptune is standard practice in ML research for experiment tracking.

pufferlib