The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Documentation in references/vendor/impeccable/impeccable.md and hooks.md explicitly describes a mechanism where a self-contained binary tool is downloaded by a launcher if it is not already present in the skill directory.
[REMOTE_CODE_EXECUTION]: The skill directs the agent to execute the downloaded external binary. The instructions in hooks.md specifically warn that the engine runs 'unsandboxed with the agent's environment and network access', which constitutes a high-risk remote code execution pattern for third-party tools.
[COMMAND_EXECUTION]: The skill contains extensive instructions to execute shell commands via the 'impeccable' CLI, including starting local servers (live-server), attaching to Chrome DevTools debugging ports (9222), and executing commands on mobile device simulators using adb and xcrun simctl.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to inspect and process live web pages, ingesting content such as element.outerHTML and textContent into the agent's context. This creates a vulnerability where a malicious web page could embed instructions intended to influence the agent's design generation or behavior, especially since boundary markers and sanitization are not specified.
[PROMPT_INJECTION]: Instructions in references/vendor/impeccable/craft.md and 'degraded' agent role files (e.g., asset-producer.md) instruct the agent to conceal certain persona substitutions or internal routing logic from the user, which can be leveraged to hide actions or intent.

ui