The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill utilizes shell execution patterns to download and run installation scripts from remote sources. Specifically, bootstrap.sh and the project's README recommend using bash <(curl ...) to execute a remote script from the author's GitHub repository. Additionally, install.sh downloads and runs the Homebrew installation script.
[EXTERNAL_DOWNLOADS]: The skill fetches several external components at runtime or during installation, including Python packages (yt-dlp, playwright, funasr, torchaudio), browser engines (Chromium), and ASR models from ModelScope. It also clones a companion repository (video-download) from the author's GitHub.
[CREDENTIALS_UNSAFE]: The skill handles WeChat session state by reading and writing to a local credential store located at ~/.workbuddy/credentials/yuanbao_state.json. Although this is used for legitimate session persistence and correctly sets restricted file permissions (chmod 600), it involves the management of sensitive session cookies.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted data from various external video platforms (YouTube, Bilibili, Douyin, WeChat Channels) to generate transcripts. This content is processed by an agent equipped with powerful tools like Bash and file system access, which could be exploited if malicious instructions are embedded in the video metadata or transcribed text.
[COMMAND_EXECUTION]: The skill extensively uses subprocess.run and subprocess.Popen to interact with system utilities such as ffmpeg, ffprobe, and yt-dlp. While it generally handles arguments as lists, the inclusion of external data (such as video titles or URLs) in these execution flows requires careful sanitization.

video-transcript