screencast-tutorial-video
Pass
Audited by Gen Agent Trust Hub on Sep 2, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads required font assets from a trusted source during the preflight stage.
- Evidence:
preflight.shusescurlto fetch Montserrat and Roboto Mono fonts from the officialgoogle/fontsrepository on GitHub (github.com/google/fonts/raw/main/ofl/...). - [COMMAND_EXECUTION]: The skill relies on several external CLI tools to perform its core functions of video recording and processing.
- Evidence:
preflight.shinstalls and verifiesffmpeg,vhs, andcliclickvia Homebrew, andplaywrightvia NPM. These tools are used throughout the scripts (e.g.,hands.sh,render-tape.sh,still-scene.sh) to capture screens, simulate input, and mux video files. This execution is standard for the skill's primary purpose. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-controlled data that could potentially contain malicious instructions, though it includes oversight mechanisms.
- Ingestion points: Reads
storyboard.mdfrom the build directory (SKILL.md, step 1) and accepts user-defined scene specifications in JSON format (browser-scene.mjs). - Boundary markers: The workflow requires an "approved" storyboard and instructs the agent to ask for clarification on
[NEEDS: ...]markers rather than fabricating values. - Capability inventory: The skill has extensive shell execution, file writing, and network access (via TTS APIs) capabilities.
- Sanitization: Inputs like the command text for command cards (
make-card.sh) are written to files to avoid shell escaping issues, reducing command injection risks. - [DYNAMIC_EXECUTION]: The skill dynamically generates and executes scripts as part of its rendering pipeline.
- Evidence:
render-tape.shgenerates temporary VHS.tapefiles by merging templates with storyboard content and executes them using thevhstool.browser-scene.mjsutilizes Playwright'spage.evaluateto execute JavaScript for scrolling, highlighting, and form interactions within a headless browser context. These behaviors are essential for the skill's automation goals.
Audit Metadata