The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bash script located at /usr/local/bin/speak to process audio. The script utilizes standard system utilities including curl, jq, and afplay (macOS audio player).
[INDIRECT_PROMPT_INJECTION]: The skill processes text content that may originate from untrusted external sources (e.g., web pages read by the agent).
Ingestion points: Text is passed as a command-line argument to the speak script.
Boundary markers: The instructions advise the agent to escape double quotes within the text argument.
Capability inventory: The script performs local network requests (curl to localhost), file system writes (/tmp/), and shell command execution.
Sanitization: The speak.sh script uses jq --arg to safely encode the input text into a JSON payload, which effectively mitigates command injection or payload manipulation within the TTS service request.
[DATA_EXPOSURE]: The skill uses temporary files in /tmp/ to store text and generated audio. These files are scoped to the current user via the ${USER} environment variable, following standard local development practices.

speak