The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines evaluation protocols that interpolate untrusted model outputs into system prompts for automated judging.
Ingestion points: Placeholders such as {response}, {response_a}, and {response_b} are used in SKILL.md to ingest external, potentially adversarial content into the evaluation context.
Boundary markers: The skill recommends using structured JSON output formats and requires chain-of-thought justification before scoring to help ground the model's judgment and isolate it from the input content.
Capability inventory: No sensitive file access, network operations, or shell execution capabilities are defined within the skill's own content.
Sanitization: The skill does not explicitly provide sanitization instructions for the interpolated strings, instead relying on instructional constraints to mitigate the impact of the input content.
[EXTERNAL_DOWNLOADS]: The skill provides links to external research papers and technical blog posts.
Evidence: References include links to academic resources on arxiv.org and industry research on eugeneyan.com. These are informational and do not involve the automated download or execution of scripts or packages.

advanced-evaluation