The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text and asset descriptions extracted from external websites during the capture phase, which can be manipulated to include malicious instructions for the agent.
Ingestion points: Website content is stored in capture/extracted/visible-text.txt and capture/extracted/asset-descriptions.md for subsequent processing by sub-agents.
Boundary markers: The instructions do not define specific delimiters or "ignore embedded instructions" warnings for the agent when it interpolates website data into its reasoning process.
Capability inventory: The skill has access to the hyperframes CLI tool, the ability to execute Node.js scripts (node), and can perform network operations via curl to external APIs.
Sanitization: The workflow incorporates a linter (npx hyperframes lint) and a custom verification script (w2h-verify.mjs) to detect and correct common build errors and inconsistencies.
[DYNAMIC_EXECUTION]: The skill programmatically generates HTML and JavaScript files to create video compositions, which are then rendered in a browser environment.
Evidence: Step 5 of the workflow involves generating index.html and multiple beat-level composition files that include GSAP-based animation logic and library imports from CDNs.

website-to-hyperframes