The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill creator instructions specifically direct the agent to "write and run a script" for grading assertions during the evaluation process. Additionally, the scripts/init_skill.py file dynamically generates Python scripts from hardcoded templates and saves them to the filesystem.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data that could contain malicious instructions.
Ingestion points: Reads and processes user feedback from feedback.json, test prompts from evals.json, and the generated outputs from other skills being tested.
Boundary markers: The instructions do not define explicit boundary markers or "ignore" instructions when the agent processes feedback or test results.
Capability inventory: The skill can spawn subagents, execute multiple local Python scripts (scripts.aggregate_benchmark, eval-viewer/generate_review.py, scripts.run_loop), write files, and package skill archives.
Sanitization: There is no evidence of sanitization or strict schema validation for the contents of the feedback or test outputs before they are used to influence subsequent skill generation or improvement steps.
[COMMAND_EXECUTION]: The SKILL.md instructions guide the agent to execute various local Python modules and system commands using shell syntax, including nohup, kill, and python -m scripts.*.
[PRIVILEGE_ESCALATION]: The scripts/init_skill.py script uses chmod(0o755) to grant execution permissions to the example.py script it generates. While typical for setup tasks, it represents a pattern of privilege modification on the local filesystem.

skill-creator