The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute multiple shell commands including npm create vite@latest, npm install, and npm run dev. It also specifically directs the agent to auto-open the browser using the open command on macOS immediately after the server starts, bypassing user confirmation for the browser launch.
[DYNAMIC_EXECUTION]: The skill generates a complete React application at runtime and immediately executes it using npm run dev. This involves running code that the agent has just generated or modified, which is a form of dynamic execution.
[EXTERNAL_DOWNLOADS]: The skill downloads and installs dependencies from the NPM registry (vite, react, tailwindcss, gsap, lucide-react) at runtime. While these are well-known packages, the use of @latest without version pinning and the automated execution of the resulting environment increases the attack surface.
[PROMPT_INJECTION]: The instructions include directives to override standard agent behaviors, such as explicitly telling the agent not to run an intake questionnaire and claiming that a specific user ('Taoufik') has authorized this deviation from normal protocols. It also contains 'Trigger even when...' instructions in the metadata to force the agent to use this specific skill over others.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided brand names, purposes, and value propositions to generate site content and code. It lacks explicit boundary markers or sanitization instructions for this user-controlled data before it is interpolated into the generated project files, creating a surface for indirect injection.

build-site