The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection (Category 8) because its primary function involves extracting content and design tokens from untrusted external sources.
Ingestion points: Data is ingested from external URLs, web captures, screenshots, and documents (PDF/DOCX) within the copy.md, inputs.md, and redesign.md reference files.
Boundary markers: The skill includes explicit instructions for the agent to discard directives, prompts, or behavioral suggestions found in the source material (e.g., HTML comments, script tags, embedded text).
Capability inventory: The skill has the ability to write files to the local project root (DESIGN.md), execute local shell commands to run a preview server, and perform network requests to fetch content.
Sanitization: No programmatic sanitization or filtering is performed on the extracted content; the skill relies entirely on the agent's instruction-following to ignore embedded directives.
[COMMAND_EXECUTION]: The skill instructs the agent to execute a local server using the Bun runtime to facilitate design previews and interaction recording.
The execution command is bun run scripts/preview-server.ts with various flags for session paths and ports.
The preview-server.ts script is included in the skill's scripts/ directory and contains logic to serve HTML fragments and record events to a local file.
The server binds only to 127.0.0.1 and includes path traversal validation to ensure it only serves files from the designated session directory.

design-builder