The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches pre-trained large language models and calibration datasets from Hugging Face's official infrastructure. These operations are essential for the quantization process and target well-known, trusted services.
[COMMAND_EXECUTION]: The instructions include standard shell commands for environment setup (e.g., pip install auto-gptq) and model deployment via Docker. These are expected administrative tasks for configuring machine learning tools.
[PRIVILEGE_ESCALATION]: The troubleshooting documentation suggests using sudo apt-get install to install system-level dependencies. While this involves elevated privileges, it is a standard practice for manual environment setup and is presented as a user-executed fix rather than an automated script behavior.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection through its ingestion of external datasets for calibration.
Ingestion points: Data enters the agent's context via the load_dataset function when fetching calibration sets like C4 or ShareGPT in SKILL.md and references/calibration.md.
Boundary markers: The instructions do not specify explicit delimiters or "ignore embedded instructions" warnings for the processed calibration data.
Capability inventory: The skill has the capability to perform file system writes (save_quantized), network operations (push_to_hub), and subprocess execution (pip install).
Sanitization: There is no mention of filtering or sanitizing the calibration text before it is used to calculate model weights.

gptq