The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines functions that ingest potentially untrusted data (documents, task seeds, and instruction examples) and interpolate them directly into prompts used for LLM-based data generation.
Ingestion points: The document parameter in generate_qa, the seeds parameter in self_instruct, and the example dictionary in format_instruction (all in SKILL.md).
Boundary markers: Absent. The templates do not use delimiters or instructions to ignore embedded commands within the data.
Capability inventory: The skill utilizes model.generate to process the interpolated prompts. No high-risk capabilities like file system writes or network operations are explicitly defined in the provided snippets.
Sanitization: No sanitization or validation of the input content is implemented before it is placed into the prompt strings.

dataset-engineering