The Agent Skills Directory

[DYNAMIC_EXECUTION]: Static analysis flagged several instances of yaml.load() as potentially unsafe. However, a manual code audit of the runtime components (capture.py, loading.py, runner.py, and shared_validate_design_system.py) confirms that all custom YAML loaders used by the skill inherit from yaml.SafeLoader, which is the industry standard for preventing arbitrary code execution during deserialization.
[COMMAND_EXECUTION]: The skill utilizes subprocess.run to interact with git and coordinate internal runtime utilities (e.g., validators and eval runners). These executions are strictly controlled, targeting either the skill's own distributed Python scripts or performing information-gathering commands like git rev-parse within user-authorized repository roots.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from external repositories, design documents, and web pages. To mitigate the risk of indirect prompt injection, it employs a deterministic capture process that validates all extracted data against rigid JSON schemas, uses strict 'closed-object' mapping, and explicitly refuses to execute source code found in the target repositories.
[DATA_EXFILTRATION]: The skill's instructions and 'Experience' guidelines explicitly prohibit the recording of credentials, secrets, or internal URLs. File system operations are confined to safe, relative paths within authorized project roots, and network operations are restricted to specific, non-malicious metadata extraction from provided URLs.

ui-template-author