The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone several repositories for implementation. Specifically, it references github.com/locuslab/wanda and github.com/IST-DASLab/sparsegpt. These are the official research implementations for the Wanda and SparseGPT techniques from established academic labs (Locus Lab and IST-DASLab) and are considered safe.
[COMMAND_EXECUTION]: The skill uses standard package management commands (pip install) and model loading commands through the transformers library. These are routine operations for the stated purpose of model compression.
[REMOTE_CODE_EXECUTION]: While the skill involves downloading external code, the sources are official research repositories associated with the specific papers referenced (ICLR 2024 and arXiv). The operations are consistent with the primary purpose of the skill and do not involve piped execution or untrusted third-party scripts.
[INDIRECT_PROMPT_INJECTION]: The skill ingests calibration data (e.g., C4 or WikiText) to calculate activation statistics. However, this data is processed numerically to calculate importance scores for weights and is not interpolated into LLM prompts in a way that could trigger instruction-following behavior. The attack surface is minimal and intended for statistical analysis.

model-pruning