The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's installation process involves cloning its own repository and subsequently downloading several third-party libraries from GitHub accounts including 'haidrrrry', 'Vincentwei1021', and 'new-xp'.
[REMOTE_CODE_EXECUTION]: The skill requires the user to execute a bash script (scripts/install.sh) downloaded from the repository. This script performs further downloads and system modifications to link skill files and install dependencies.
[COMMAND_EXECUTION]: The instructions mandate the use of shell commands like ffmpeg to verify video renders and npx for installation, which involves spawning subprocesses with potentially complex arguments.
[INDIRECT_PROMPT_INJECTION]: The skill incorporates tools for app walkthroughs and web page captures (Puppeteer), which ingest untrusted data from the internet.
Ingestion points: Web page text, application screenshots, and scraped metadata in SKILL.md.
Boundary markers: The skill contains a manual instruction (Rule 4) telling the agent to treat fetched content as data and not instructions, but lacks structural delimiters.
Capability inventory: The skill possesses extensive capabilities including file system writes, shell execution (ffmpeg), and network access via capture tools.
Sanitization: There is no evidence of programmatic sanitization; the skill relies on the agent's ability to 'flag' suspicious content manually.

aegntic-video