The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to how it handles external data during the video generation process.
Ingestion points: The skill fetches untrusted text from the App Store (metadata and descriptions) via fetch_appstore_screens and website content via capture_website.
Boundary markers: The prompt templates provided in the reference files do not utilize XML tags, triple backticks, or other delimiters to isolate this untrusted content from the system instructions.
Capability inventory: The skill can trigger expensive, paid video generation calls through providers like Seedance and Kling.
Sanitization: There is no explicit sanitization or filtering logic applied to the external text before it is interpolated into model prompts.
[SAFE]: The skill implements a mandatory cost transparency gate (identity_balance), which requires explicit user consent before proceeding with any operation estimated to cost more than $5, protecting against unintended financial expenditure.
[SAFE]: A wall-clock guard is enforced at runtime to prevent resource exhaustion by terminating the process if the first generation call is not initiated within 5 minutes of asset sourcing.
[SAFE]: The skill employs an 'Avatar-type probe' using analyze_media to detect and block the use of trademarked or copyrighted characters in generated content, reducing legal and moderation risks.

app-sizzle