The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements patterns to process untrusted text data through an LLM, which inherently presents an indirect prompt injection surface. The analysis found the following controls:
Ingestion points: Untrusted data enters via the text parameter in extraction functions like extractContactInfo and extractReview (SKILL.md).
Boundary markers: The functions utilize system messages to define the extraction task, though they lack specific instructions for the model to disregard instructions contained within the user text.
Capability inventory: The skill is limited to calling the OpenAI API; it does not perform file writes, subprocess execution, or arbitrary network operations.
Sanitization: The implementation uses Zod schemas to perform strict runtime validation (e.g., ProductReviewSchema.parse(data)) on the LLM's output, which effectively ensures the data structure conforms to expectations even if an injection occurs.
[SAFE]: The skill utilizes well-known, official libraries including openai and zod. All code snippets represent industry-standard integration patterns. No evidence of obfuscation, hardcoded credentials, or persistence mechanisms was found.

structured-output-extractor