The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements 'Model-Based Graders' which use an LLM to evaluate open-ended outputs, code changes, or task completion. This creates a potential surface for indirect prompt injection where the data being graded (such as code outputs or external task results) could contain instructions designed to manipulate the grader's scoring or logic.
Ingestion points: Evaluation definitions and logs are stored in the .claude/evals/ directory, and the skill processes arbitrary code outputs during the grading phase of the workflow.
Boundary markers: While the skill uses markdown headers and tags (e.g., [MODEL GRADER PROMPT]) to structure evaluation logic, it does not explicitly define boundaries or instructions to ignore commands embedded within the data being evaluated.
Capability inventory: The skill has access to the Bash, Read, Write, and Edit tools, allowing it to perform file operations and execute shell commands based on evaluation outcomes.
Sanitization: No evidence of sanitization or escaping of external content is present before it is interpolated into the grading prompts.

eval-harness