The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an attack surface for Indirect Prompt Injection (Category 8) due to its ingestion of untrusted external data. Ingestion points: In Phase 1 Step 1, the skill reads reference files provided by the user, and in Phase 5 Step 3-1, it accepts external URLs and images for design analysis. Boundary markers: The instructions lack explicit delimiters or 'ignore' commands to prevent the AI from being influenced by instructions embedded within the user-provided files or analyzed web pages. Capability inventory: The skill has extensive file system write access (creating numerous project documents), network navigation capabilities via Playwright, and the ability to orchestrate multiple sub-agents with specialized roles. Sanitization: There is no evidence of validation, filtering, or sanitization of content retrieved from external sources before it is processed by the agents.
[COMMAND_EXECUTION]: The skill utilizes Playwright tools to perform automated browser interactions with user-specified URLs in Phase 5. Navigating to and analyzing the accessibility and visual structure of untrusted remote websites presents a security risk if those sites contain malicious payloads designed to exploit browser automation tools or the agent's interpretation logic.

plan