The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it is designed to process and evaluate untrusted LLM-generated responses.
Ingestion points: Untrusted content is ingested through the dataset parameter in the GradingRunner.arun() method (SKILL.md, pipeline.md) and in the various analyzer analyze() methods (analyzer.md).
Boundary markers: The documentation does not specify the use of delimiters or explicit instructions to the judge model to ignore embedded commands within the evaluation samples.
Capability inventory: The framework utilizes LLMGrader for rubric-based prompting and AgenticGrader for autonomous code execution and verification.
Sanitization: There is no evidence of input sanitization or escaping of the response or query fields before they are interpolated into the evaluation prompts.
[DYNAMIC_EXECUTION]: The skill performs runtime code execution to facilitate automated grading. The AgenticGrader class (graders.md) executes code snippets provided in Checkpoint.content to verify grading criteria, and the CodeExecutionGrader (graders.md) runs candidate code to calculate test pass rates. These capabilities are primary to the skill's purpose and are documented as using isolated sandboxes.
[COMMAND_EXECUTION]: The AgenticGrader (graders.md) shells out to external command-line interfaces such as claude, codex, and cursor to perform evaluations using external agents.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the py-openjudge package from PyPI and relies on the presence of external CLI-based agent tools (e.g., Anthropic's Claude Code) for its agentic grading functionality.

01-graders-and-pipeline