The Agent Skills Directory

[PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it ingests untrusted external data—such as benchmarks, rubrics, and task YAML/JSON files—to generate executable grading logic.
Ingestion points: Benchmark source files, task.yaml, task.json, and rubric text are read from the target directory as specified in SKILL.md.
Boundary markers: Absent. The instructions do not provide explicit boundary markers or delimiters to separate untrusted user data from the generated code.
Capability inventory: The skill generates executable .py and .sh files and runs shell commands via the skillevaluator CLI.
Sanitization: Absent. There is no guidance for validating or sanitizing user-provided rubric content before it is incorporated into the generated scoring logic.
[COMMAND_EXECUTION]: The skill facilitates dynamic code execution by instructing the agent to scaffold and implement grader.py or grader.sh scripts, which are subsequently executed by the skillevaluator tool in a verifier context. This represents the generation of executable content from user-provided templates.

create-custom-grader