The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection. It instructs the agent to ingest untrusted data from the user (SKILL.md, Step 1), including product descriptions, specific text beats, and attached files such as screenshots, results, and brand assets. This content is then interpolated into a build prompt (SKILL.md, Step 3) to generate executable HTML, Javascript, or HyperFrames composition code.
Ingestion points: Step 1 collects user-supplied text beats, proof files (screenshots/results), and brand assets.
Boundary markers: There are no explicit delimiters or instructions to treat user-provided content as data only, lacking "ignore embedded instructions" warnings.
Capability inventory: The skill uses the HyperFrames CLI for validation and rendering and generates code that includes dynamic execution capabilities like window.seek(seconds).
Sanitization: No sanitization, escaping, or validation of the input content is mentioned before it is processed into the code generation phase.
[EXTERNAL_DOWNLOADS]: The skill mandates the use of an external production runtime and CLI called HyperFrames (SKILL.md, Step 5). It specifically instructs the agent to "search its registry before hand-building reusable effects," which implies fetching and incorporating code or assets from a remote, third-party repository that is not among the recognized trusted providers.
[COMMAND_EXECUTION]: The skill workflow requires the execution of the HyperFrames CLI for linting, checking, previewing, and rendering the final video (SKILL.md, Step 5). These commands are executed within the agent's environment on a project structure that has been populated with untrusted user inputs and potentially third-party registry assets.

launch-video