The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates two JavaScript files, _copy_screenshot.mjs and _copy_auth.mjs, and executes them using the node runtime. It also uses shell commands like find, grep, and curl to inspect the project structure and verify the status of the development server.
[PROMPT_INJECTION]: Phase 3c instructs the agent that rules should be "applied automatically — no permission needed." This encourages the agent to bypass human-in-the-loop review, potentially leading to unauthorized or incorrect modifications to the project's source code.
[CREDENTIALS_UNSAFE]: The authentication script (_copy_auth.mjs) is designed to capture and use sensitive environment variables (TEST_PASSWORD) to log into the application. The resulting authentication state is saved to /tmp/copy-auth-state.json. Files stored in /tmp are often globally readable on multi-user systems, creating a risk of session hijacking or credential exposure.
[PROMPT_INJECTION]: The skill performs a comprehensive audit of all user-facing strings (Phase 1b), which creates an indirect prompt injection attack surface where malicious instructions embedded in the application's UI or internationalization files could influence the agent's behavior.
Ingestion points: UI components (.tsx, .jsx) and localization files (.json) found within the project's source directory.
Boundary markers: None. The skill does not provide the agent with instructions to ignore or delimit embedded commands within the extracted UI text.
Capability inventory: File system modification, shell command execution, and Node.js script execution.
Sanitization: None. The extracted strings are processed directly for rewriting based on provided principles.

copy