The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The scripts/setup.sh script clones the whisper.cpp repository from GitHub (https://github.com/ggerganov/whisper.cpp.git) and downloads a large-v3 model file (~3GB). This process also involves building the whisper-cli binary locally using cmake and make.\n- [COMMAND_EXECUTION]: The skill executes shell commands to process data and manage dependencies.\n
It uses subprocess.run in Python to invoke a bash helper script (transcribe-audio.sh).\n
The shell scripts use ffmpeg for audio conversion and the whisper-cli for transcription to handle voice messages.\n
These executions use list-based arguments and quoted variables, which are safe practices to prevent command injection from user-supplied filenames found in exports.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted data from WhatsApp exports.\n
Ingestion points: Untrusted data enters the agent context through the parsing of _chat.txt and exported media filenames within the process-conversation.py script.\n
Boundary markers: The skill formats messages into Markdown chunks but does not include explicit delimiters or instructions to tell the agent to ignore instructions embedded within the chat text.\n
Capability inventory: The skill has the capability to read and write files in the local project directory and execute subprocesses like ffmpeg and the whisper-cli.\n
Sanitization: The scripts perform regex-based removal of zero-width control characters but do not sanitize message content for potential prompt injection strings.

read-whatsapp-export