The Agent Skills Directory

[COMMAND_EXECUTION]: The prepare.cs script executes the ffmpeg system utility as a subprocess via ProcessStartInfo to perform video frame extraction and image similarity calculations. Spawning shell processes with variable arguments (such as user-provided output paths) is a sensitive operation, although the implementation uses non-shell execution for improved safety.- [EXTERNAL_DOWNLOADS]: The skill performs network requests to api.bilibili.com and Bilibili's content delivery networks to download video metadata and binary files. These operations target well-known service domains required for the skill's stated purpose and are documented neutrally.- [INDIRECT_PROMPT_INJECTION]: The skill's workflow involves transcribing all text and code found within video frames using AI. This creates a surface for indirect prompt injection where malicious instructions visually presented in the video (e.g., text stating 'Ignore all prior rules') could influence the agent's behavior when the transcription is incorporated into the prompt context.
- Ingestion points: Images extracted from Bilibili videos (SKILL.md, Step 2).
- Boundary markers: The workflow uses markdown headers to organize frames but lacks explicit delimitation or instructions to treat embedded visual text as untrusted data.
- Capability inventory: Subprocess execution of ffmpeg, local file system writes, and network access.
- Sanitization: None; the AI is instructed to transcribe 'all text/code' without any filtering or validation.- [METADATA_POISONING]: There is a discrepancy between the provided source code (prepare.cs in C#) and the documentation files (installation.md and examples/basic-usage.md), which describe a Python-based implementation using yt-dlp and requests. This misleading metadata can lead to confusion regarding the skill's true dependencies and execution environment.

bilibili-analyzer