The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the user or agent to execute shell commands using xcrun simctl and xcodebuild for simulator management and ffmpeg for generating tiled filmstrips from screen recordings. These commands are standard for iOS development and are used legitimately here to provide evidence for the design review.
[EXTERNAL_DOWNLOADS]: The skill references external developer tools, specifically ffmpeg and rocketsim, which are required to fulfill the mandatory evidence capture protocol. While these are well-known tools, they represent external dependencies.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted user-supplied SwiftUI code and descriptions to generate its review verdicts.
Ingestion points: User-provided code and descriptions are directly interpolated into the AI subagent's prompt template in references/reviewer-prompt.md.
Boundary markers: The prompt template lacks explicit delimiters or instructions to the subagent to ignore instructions potentially embedded within the code being analyzed.
Capability inventory: The skill performs shell-based evidence capture (e.g., xcrun simctl io booted screenshot) and file system access to the repo root and asset catalogs.
Sanitization: No validation or sanitization is performed on the reviewed content before it is processed by the AI subagent.

adversarial-ios-design