The Agent Skills Directory

[SAFE]: The skill does not contain any detected prompt injections, obfuscation, or malicious persistence mechanisms. It is authored by a trusted vendor and performs actions consistent with its documented purpose.
[EXTERNAL_DOWNLOADS]: The skill downloads model weights (checkpoints) from a public cloud storage bucket at rfd4-proteina-public-e04a.cwobject.com. It also installs the AtomWorks dependency via pip from a pinned Git revision specified in the model's pyproject.toml. These downloads are documented as necessary for the model's operation.
[COMMAND_EXECUTION]: Shell commands such as nvidia-smi, make setup, and rfd4-generate are used to verify hardware availability and perform the installation. These are standard operations for setting up a machine learning environment and are executed within the user-defined RFPROTEINA_ROOT.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from local configuration files (configs/inference.yaml, .env, and .env.local). Although these represent an attack surface for indirect prompt injection, the skill mitigates this risk by utilizing structured parsing (YAML) and providing explicit instructions to the agent to treat these files as data and never shell-source them.

rfd4-proteina-setup