The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media files (audio/video) that could contain malicious instructions designed to manipulate the agent.
Ingestion points: Media bytes, filenames, and generated transcripts processed via shell tools.
Boundary markers: Includes a dedicated 'Untrusted content boundary' section with explicit instructions to ignore transcribed requests and use delimiters when passing text to the agent.
Capability inventory: Shell command execution via ffmpeg and whisper-cli, and file system writes for transcripts and sidecars.
Sanitization: Instructions mandate unprivileged sandboxing, path resolution under a project root, and avoiding shell interpolation by passing paths as argv entries.
[COMMAND_EXECUTION]: The skill utilizes shell commands for audio normalization and transcription. It mitigates risk by instructing the agent to use unprivileged processes, fixed parameter sets for reproducibility, and sandboxing with network access disabled.
[DYNAMIC_EXECUTION]: An embedded Python script is provided to verify the SHA-256 digests of the transcription engine and model file against a locally reviewed whisper-artifacts.json manifest. This mechanism ensures that the agent only uses verified binaries and weights, protecting against the execution of unauthorized or tampered code.

video-transcribe