The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill templates in SKILL.md, references/examples.md, and scripts/evaluation_example.py interpolate untrusted data (user prompts and model responses) into judge prompts.
Ingestion points: The {prompt} and {response} variables in the evaluation templates are used to inject content for evaluation.
Boundary markers: The templates use Markdown headers (e.g., ## Response to Evaluate) to structure the prompt, but they lack specific instructions to ignore embedded commands or system prompt overrides within the untrusted content.
Capability inventory: The skill is primarily instructional. The provided Python examples in scripts/evaluation_example.py are limited to safe logic and printing, with no dangerous capabilities such as file system writing, network access, or shell execution.
Sanitization: The provided examples and templates do not demonstrate any sanitization, escaping, or filtering of the input data before interpolation.

advanced-evaluation