The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a framework where external data, such as agent outputs and user-defined test inputs, are processed by a model-based grader.
Ingestion points: Untrusted content enters the system via the input and actual output variables used in the grading prompt templates defined in SKILL.md.
Boundary markers: The template uses basic structural labels (e.g., INPUT: {input}) but does not employ hardened delimiters or specific instructions to the grading model to disregard potentially malicious instructions within the data.
Capability inventory: The skill is configured to use Bash, Write, and Edit tools to create and run scripts.
Sanitization: The instructions do not detail any sanitization or escaping logic for data before it is interpolated into the grader prompt.
[COMMAND_EXECUTION]: The skill's operational procedure requires the agent to generate and execute Bash scripts (e.g., evals/run-agent-routing.sh) to automate the evaluation suite execution.
[DYNAMIC_EXECUTION]: The workflow relies on the dynamic creation of executable shell scripts and YAML configuration files to define the evaluation logic at runtime.
[SAFE]: The skill incorporates advanced statistical techniques for result quantification, referencing research and best practices from organizations such as Anthropic and OpenAI.

eval-harness