The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from mockup files located in artifacts/mockup-sandbox. If these files are compromised or contain malicious instructions disguised as code, the agent might process and integrate them into the production codebase.
Ingestion points: Reads component files from artifacts/mockup-sandbox in Step 1 to extract design decisions and imports.
Boundary markers: The instructions do not specify any delimiters or safety warnings to distinguish between the code being analyzed and instructions to the agent.
Capability inventory: The skill has the ability to write new files to the main app, install Node.js packages via pnpm, and modify existing routing configurations.
Sanitization: There are no explicit validation or sanitization steps mentioned to verify the safety or integrity of the mockup code before it is transformed and moved.
[EXTERNAL_DOWNLOADS]: Step 4 instructs the agent to install missing dependencies using pnpm --filter @workspace/<target-slug> add <package>. Because the package names are determined dynamically from the imports within the mockup, an attacker who can influence the mockup content could trigger the installation of malicious or typosquatted packages into the production environment.
[COMMAND_EXECUTION]: The skill uses shell commands for package management and environment setup. While these are intended for the skill's primary function of graduating code to production, the reliance on shell execution for dynamically identified packages introduces a vector for command-related issues if inputs are not carefully handled by the agent.

mockup-graduate