The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several internal Node.js scripts (tokenize-css.js, apply-overrides.js, gallery.js) to process site files, measured design system tokens, and user configuration. These scripts have read and write access to the local project directory to transform existing styles and assets.
[DYNAMIC_EXECUTION]: The apply-overrides.js script dynamically generates JavaScript and TypeScript files, such as remix-loader.ts and patch.<slug>.js. These files contain logic for overriding site behaviors and structure, and are imported at runtime by the target application using Vite's import.meta.glob mechanism. This constitutes a medium-level risk associated with runtime script generation and execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources including TEARDOWN.md, design token JSON files from cloned sites, and user-supplied copy. This untrusted content is used to populate design templates and generates briefs for external asset generation services.
Ingestion points: TEARDOWN.md, tokens-*.json, motion.json, and user-provided brand copy/URLs.
Boundary markers: Absent; the instructions do not define explicit delimiters or warnings to prevent the agent from following instructions embedded within the ingested site data.
Capability inventory: File system read/write, Node.js process execution, and browser-side script injection via generated patches.
Sanitization: Basic sanitization is applied to direction slugs, and regex escaping is used during CSS tokenization, but no comprehensive sanitization of text nodes or metadata is present.
[EXTERNAL_DOWNLOADS]: The skill instructions specify using Higgsfield, an external AI service, to generate hero and decorative assets. This involves sending composition briefs to an external service and downloading generated image files into the site's public directory structure.

remix-site