The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md direct the agent to execute generated prototypes using project task runners such as pnpm, python, and bun. This involves running local shell commands based on generated content.
[DATA_EXFILTRATION]: LOGIC.md suggests creating self-contained HTML files intended to be shared via email or other portable means. This poses a potential risk of inadvertent data exposure if the agent embeds sensitive business logic or internal state data into these shareable files.
[PROMPT_INJECTION]: The skill processes untrusted input from user prompts to generate executable prototype code, representing an indirect prompt injection surface.
Ingestion points: User descriptions of business logic and existing project source code (found in SKILL.md).
Boundary markers: None specified in the instructions to prevent instructions within the input data from influencing code generation.
Capability inventory: Creation of executable HTML files and application routes, and execution of shell commands via pnpm, python, and bun (found in SKILL.md, LOGIC.md, and UI.md).
Sanitization: The skill lacks instructions for sanitizing or escaping user-provided data before it is interpolated into the generated code.
[SAFE]: The skill includes security best practices by instructing the agent to gate prototype switcher UI behind environment checks (e.g., process.env.NODE_ENV !== 'production') to ensure debug tools are not exposed in production builds.
[SAFE]: The instructions recommend using isolated 'scratch' databases or local files for prototype persistence, preventing interference with existing data sets.

prototype