The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes transcription data that may originate from untrusted audio sources. This data is then formatted into a Markdown report and passed to a vision model (Gemini) for analysis. This workflow is susceptible to indirect prompt injection where instructions hidden in the audio or transcript could influence the AI's behavior or final output.
Ingestion points: The skill reads transcription data from JSON files provided by the user in scripts/fix_hallucinations.py and scripts/enrich_uncertain_with_vision.py.
Boundary markers: There are no explicit instructions or delimiters used to separate the untrusted transcription text from the agent's instructions when performing vision-based review in the documented workflow.
Capability inventory: The skill possesses the capability to execute shell commands (ffmpeg, whisper-cli) and write files to the local filesystem (extracted frames and reports).
Sanitization: The enrich_uncertain_with_vision.py script performs minimal escaping of special characters (only the pipe | character) when embedding the transcript text into the generated Markdown report.
[COMMAND_EXECUTION]: The skill's scripts execute external binaries to process audio and video data.
Evidence: subprocess.run is used in scripts/fix_hallucinations.py and scripts/enrich_uncertain_with_vision.py to call ffmpeg and whisper-cli. While the use of list-based arguments prevents typical shell injection attacks, the skill depends on the presence and integrity of these local binaries.

whisper-corruption-recovery