The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data by capturing content from user-provided URLs using npx hyperframes capture. This data is extracted into files like visible-text.txt and asset-descriptions.md, which the agent then reads to formulate video strategies and scripts. A malicious website could embed instructions to manipulate the agent's creative output or attempt to leak session information.
Ingestion points: capture/extracted/visible-text.txt and capture/extracted/asset-descriptions.md are read in Steps 2 and 3.
Boundary markers: The instructions do not define clear delimiters for the external content before processing.
Capability inventory: The skill can execute shell commands (npx, curl), write files, and spawn sub-agents.
Sanitization: No explicit sanitization or filtering of the captured web text is performed before it is used in the prompt context.
[COMMAND_EXECUTION]: The skill relies on various shell commands to perform its core functions, including npx hyperframes for capturing, rendering, and linting, as well as curl for interacting with external Text-to-Speech (TTS) APIs. While these are necessary for the skill's utility, they represent an execution surface for the agent.
[EXTERNAL_DOWNLOADS]: The skill interacts with external services including Google Gemini (for image descriptions), ElevenLabs, and HeyGen (for TTS). It also uses npx to run the hyperframes toolkit, which is an external dependency managed by the vendor.
[DYNAMIC_EXECUTION]: In Step 5, the agent and its sub-agents dynamically generate HTML and JavaScript (GSAP) files to create video compositions. These files are subsequently executed in a headless browser environment for validation, linting, and snapshotting via the hyperframes CLI tools.

website-to-video