The Agent Skills Directory

[PROMPT_INJECTION]: The skill operates by visiting arbitrary, user-provided URLs and extracting content using a browser MCP. This creates an attack surface for indirect prompt injection, as malicious instructions hidden within a target page's DOM, text, or metadata could attempt to manipulate the agent's logic during the spec-writing or component-building phases.
Ingestion points: Browser MCP (Playwright/Puppeteer/Chrome) reading arbitrary web content.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the extraction workflow.
Capability inventory: File system access (writing specs and code), network operations (downloading assets, calling Higgsfield API), and command execution (running local scripts and Playwright tests).
Sanitization: The instructions do not specify sanitization or validation of the extracted content before it is used to generate spec files or code components.
[COMMAND_EXECUTION]: The skill executes local JavaScript scripts (tokens-probe.js and motion-probe.js) located in a sibling directory (../clone-site/scripts/) to analyze design tokens and motion parameters. It also utilizes Playwright for automated pixel diffing (toHaveScreenshot()) to verify the fidelity of the generated clone.
[EXTERNAL_DOWNLOADS]: As part of its 'asset resolution ladder,' the skill is instructed to download original media files (images, videos, fonts) directly from the target website's host to the local project directory. It also interfaces with the external Higgsfield MCP for asset generation when original files cannot be retrieved.

dom-clone