The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Python subprocess module to execute system-level commands. Specifically, scripts/run_eval.py and scripts/improve_description.py invoke the claude CLI tool, while eval-viewer/generate_review.py uses lsof and kill for network port management.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data—such as eval prompts and feedback—and interpolates it into instructions for specialized subagents (Grader, Analyzer, and Comparator). This creates a vulnerability surface for indirect prompt injection attacks.
Ingestion points: User-provided inputs in evals.json and feedback.json.
Boundary markers: The skill lacks consistent delimiters or 'ignore' instructions when interpolating external content.
Capability inventory: Includes file system writes, subprocess execution, and spawning of subagents.
Sanitization: There is no evidence of validation or escaping of external content before it is processed by the subagents.
[PERSISTENCE]: In scripts/run_eval.py, the skill writes temporary markdown files to the .claude/commands/ directory. Although it attempts to delete these files after testing trigger rates, an interrupted process could leave persistent skill definitions in the agent's configuration environment.
[DYNAMIC_EXECUTION]: The skill dynamically generates markdown command files and subagent prompts at runtime, which are then interpreted or executed by the agent environment, particularly during the trigger optimization loop.

skill-creator