The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions advise the agent to suggest shell commands (such as pnpm, python, or bun) to run generated prototypes. Users should verify these commands before execution.
[INDIRECT_PROMPT_INJECTION]: The skill uses user-provided prompts and existing project context to generate code for Logic and UI prototypes. This represents an attack surface where a malicious prompt could influence the agent to generate harmful scripts or UI elements that are then executed or viewed by the user.
Ingestion points: Project code context and user prompts are used to determine prototype functionality and structure (described in SKILL.md and LOGIC.md).
Boundary markers: No specific delimiter or isolation instructions are provided to separate user-provided descriptions from the generated code logic.
Capability inventory: Generated code can execute arbitrary JavaScript in the user's browser or run as local scripts with file system access (via the suggested shell commands).
Sanitization: The prototype template in references/picker.md uses innerHTML for rendering content, which lacks built-in protection against cross-site scripting (XSS) if the generated content is derived from untrusted input.
[DYNAMIC_EXECUTION]: The provided reference implementation for the prototype picker in references/picker.md uses innerHTML and requestAnimationFrame to dynamically render components. While standard for rapid prototyping, innerHTML is a known vector for script injection if the content is not strictly controlled.

prototype