The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses shell commands for project setup, including npx create-next-app, pnpm install, and file management commands like mkdir, mv, and shopt. It also includes a command to approve binary builds for embedded-postgres via pnpm approve-builds, bypassing default security restrictions.\n- [EXTERNAL_DOWNLOADS]: The skill downloads and installs various Node.js packages from standard registries, including core frameworks (Next.js), ORMs (Drizzle), and payment integrations (Polar, Stripe).\n- [REMOTE_CODE_EXECUTION]: The skill instructs the agent to download and add a separate skill from an untrusted source: npx skills add https://github.com/Leonxlnx/taste-skill. This constitutes the execution of remote instructions from an unverified third party.\n- [CREDENTIALS_UNSAFE]: The agent is tasked with generating and configuring sensitive environment variables, including BETTER_AUTH_SECRET, OPENROUTER_API_KEY, and payment provider tokens, storing them in a local .env file.\n- [PROMPT_INJECTION]: The design extraction feature in references/design.md requires the agent to visit user-provided URLs and execute JavaScript to extract computed styles. This presents a vector for Indirect Prompt Injection, where content from a malicious external site could influence the agent's subsequent code generation tasks.\n
Ingestion points: User-provided URL during the project interview phase.\n
Boundary markers: None identified; the agent is not instructed to isolate or sanitize the data retrieved from the external URL.\n
Capability inventory: The agent has full file-writing, command execution, and code generation capabilities within the project directory.\n
Sanitization: The skill does not implement validation or filtering for the style data and screenshots retrieved from external websites.

start-an-app