The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes methods for evaluating LLM outputs by interpolating untrusted data (e.g., {response}, {prompt}) into evaluation prompt templates. This ingestion of external data creates a surface where the content being judged could attempt to override the evaluator's instructions.
Ingestion points: Multiple prompt templates in SKILL.md for direct scoring and pairwise comparison.
Boundary markers: The templates use Markdown headers as delimiters for data sections. The skill specifically instructs the agent to ignore length and position biases and focus on defined criteria.
Capability inventory: The skill contains no scripts, subprocess executions, file system access, or network capabilities.
Sanitization: The skill advocates for robust sanitization practices, including evidence-based justifications grounded in observable text, calibration against human-labeled datasets, and explicit instructions to ignore adversarial features like length or tone.
[NO_CODE]: The provided skill consists solely of instructional documentation in Markdown format. It does not include any executable scripts, binaries, or automated tool configurations, eliminating the risk of direct malicious code execution or privilege escalation.

advanced-evaluation