The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing standard machine learning dependencies including torch, pytorch-lightning, deepspeed, wandb, ninja, and rwkv from well-known official sources such as the PyTorch wheel index and public package registries.
[INDIRECT_PROMPT_INJECTION]: The skill includes code for processing document chunks and user-provided prompts. Ingestion points: Untrusted text data enters the agent context through the load_document() function and tokenized input sequences passed to model.forward() in SKILL.md. Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands within the processed data. Capability inventory: The skill utilizes pip install for package management and sets environment variables to trigger JIT compilation of CUDA kernels using the ninja build system. Sanitization: No input sanitization, filtering, or escaping of external document content is implemented prior to model processing.

rwkv-architecture