The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the uv package manager to execute a local script located at a specific user path (/Users/darin/clawd/skills/voice-transcribe/transcribe). This command is used to process audio files and is governed by the skill's security policies.
[DATA_EXFILTRATION]: The skill incorporates extensive behavioral instructions designed to prevent the exfiltration of transcribed text. It explicitly directs the agent to refuse any requests to pipe transcription output to network utilities such as nc, curl, or scp, identifying these as potential exfiltration vectors for private conversation data.
[CREDENTIALS_UNSAFE]: Setup instructions guide the user to store their OpenAI API key in a .env file. To mitigate the risk of key theft, the skill includes a critical guardrail that forbids the agent from reading or displaying the contents of this environment file, even under the pretext of debugging.
[PROMPT_INJECTION]: The skill identifies a surface for indirect prompt injection where untrusted content from transcribed audio files is processed by the agent.
Ingestion points: Local audio files in formats including mp3, ogg, and m4a are ingested and converted to text.
Boundary markers: The skill relies on natural language instructions and security guardrails in SKILL.md to define operational boundaries, though it does not specify technical data delimiters.
Capability inventory: The agent has the ability to execute the transcription script and utilize its output to generate responses.
Sanitization: The skill documentation does not define programmatic sanitization of the transcriptions beyond user-defined text replacements.

voice-transcribe-hardened