The Agent Skills Directory

[COMMAND_EXECUTION]: Several Python utilities in the suite (run_eval.py, improve_description.py, validate-skill.py) use the subprocess module to execute commands. These primarily interact with the local claude CLI and internal validation scripts to facilitate automated skill testing and iteration.
[EXTERNAL_DOWNLOADS]: The viewer.html template in the eval-viewer directory fetches the xlsx.full.min.js library from cdn.sheetjs.com. This well-known service is used to render spreadsheet outputs within the local interactive review interface.
[INDIRECT_PROMPT_INJECTION]: The skill creator workflow ingests untrusted user input to generate new skill instructions and descriptions, which constitutes a potential vulnerability surface.
Ingestion points: User responses gathered via the AskUserQuestion tool and JSON evaluation scenarios defined in evals/evals.json.
Boundary markers: The system does not use explicit boundary markers or delimiters when interpolating user-provided text into the generated SKILL.md files.
Capability inventory: The skill possesses Write and Edit permissions, along with the ability to execute Python scripts and trigger the claude CLI via the Bash tool.
Sanitization: There is no evidence of input sanitization or escaping for the user-provided instruction text before it is written to the file system.
[PRIVILEGE_ESCALATION]: The skill uses the Bash(chmod:*) tool to grant execution permissions to newly generated scripts (e.g., chmod +x scripts/*.py). This is standard behavior for development tools creating local automation scripts.

skill-creator