The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes numerous examples for extracting structured data from untrusted sources such as job applications, invoices, and user messages. This constitutes an attack surface where hidden instructions in the input text could influence the model.
Ingestion points: Functions like batch_extract, safe_extract, and process_csv in references/examples.md accept arbitrary text strings as input for LLM processing.
Boundary markers: The provided code snippets do not implement delimiters or explicit system instructions to treat the ingested text as untrusted content.
Capability inventory: The skill is designed to interact with various local and API-based LLM backends to generate structured output based on the provided schemas.
Sanitization: No evidence of input sanitization or filtering was found in the provided documentation or examples.\n- [EXTERNAL_DOWNLOADS]: The documentation provides standard installation instructions for well-known and reputable libraries, including outlines, transformers, and vllm, which are downloaded from official package registries.\n- [REMOTE_CODE_EXECUTION]: Configuration examples in references/backends.md for vLLM and Transformers backends include the trust_remote_code=True parameter. This setting allows the execution of code from model repositories on Hugging Face. In these examples, the parameter is used with models from established and trusted organizations.

outlines