The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, including gameplay footage, YouTube/Bilibili metadata, and user-provided or auto-generated subtitles, to generate design insights. This creates a surface for indirect prompt injection if the source content or its metadata contains adversarial instructions intended to influence the agent's behavior.
Ingestion points: The skill accepts URLs for external video platforms and local media files as described in README.md and SKILL.md.
Boundary markers: The instructions in SKILL.md attempt to mitigate accidental obedience by requiring a strict separation between observed facts, reported data, and agent inferences.
Capability inventory: The agent is authorized to download media, invoke external AI classifiers (Jev), and execute video processing tools (ffmpeg).
Sanitization: While the skill mentions blurring private data and omitting copyrighted material, it lacks specific instructions for sanitizing or escaping textual input from subtitles or commentary before interpolation into prompts.
[EXTERNAL_DOWNLOADS]: The skill fetches media from well-known video hosting services (YouTube, Bilibili) and interacts with the Vercel AI Gateway (typesafe-ai/jev) for text-based event classification.
Evidence: SKILL.md and SOURCE.md explicitly reference the use of typesafe-ai/jev at vercel.com.
Analysis: These interactions are documented neutrally as they target well-known, established services necessary for the skill's stated purpose.
[COMMAND_EXECUTION]: The skill requires the use of ffmpeg and other local media decoders to process video and audio streams.
Evidence: SKILL.md instructs the agent to use ffmpeg-derived metadata, thumbnails, and spectrograms for analysis.
Analysis: This involves executing shell commands to operate standard utility binaries typically expected in environments supporting media analysis.

worldview-gameplay-video-analysis