The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to build evaluation pipelines that interpolate untrusted data (such as user queries, model responses, and conversation context) directly into prompts for LLM-based graders. This creates a vulnerability surface where an attacker could provide malicious content in the dataset to manipulate the judge's evaluation or attempt to extract information from the evaluation context.
Ingestion points: Untrusted data enters the context via the dataset items processed by the GradingRunner in SKILL.md.
Boundary markers: While the provided prompt templates use headers like 'Context:' and 'Response:', they lack robust delimiters or specific 'ignore instructions' guards to separate untrusted content from the judge's instructions.
Capability inventory: The skill uses GradingRunner to execute Python evaluation logic and LLMGrader to perform LLM calls based on the generated templates.
Sanitization: There is no mention of sanitizing or escaping the data fields before they are interpolated into the prompt templates.
[EXTERNAL_DOWNLOADS]: The skill guides the user to install the py-openjudge package via pip to facilitate the metric design and evaluation pipeline execution.

metric-design