The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/render.mjs script utilizes child_process.spawn to execute a Chromium-based browser (Chrome, Chromium, Edge, or Brave). This execution is the core mechanism used to render the screenshot into the final marketing image. The script also accepts a CHROME_PATH environment variable to define the binary path.
[DYNAMIC_EXECUTION]: The skill dynamically generates executable content by populating an HTML template (assets/template.html) with data from a JSON configuration file. It further injects raw HTML/JavaScript fragments (scenes) into this template before rendering the page in a headless browser. The generated pages are executed with the --allow-file-access-from-files flag enabled, which is required for loading local screenshots but increases the execution surface.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied configuration data and external files that are interpolated into the rendering context.
Ingestion points: Untrusted data enters the context via the set.json configuration file, which specifies screenshot paths, headlines, and external HTML/JS scene fragments (e.g., examples/scenes/berries.html).
Boundary markers: The instructions include guidelines for the agent to maintain UI fidelity, but no technical delimiters or boundary markers are implemented in the configuration processing to separate instructions from data.
Capability inventory: The render.mjs script possesses capabilities to read local files (fs.readFileSync), write temporary files (fs.writeFileSync), and execute a web browser via shell spawning.
Sanitization: The renderer implements an escapeHtml function for textual inputs like headlines and subheadlines. However, the HTML scene fragments are read from disk and injected into the template without sanitization or validation, allowing for arbitrary JavaScript execution in the browser context.

app-store-screenshot