The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (audio files and URLs) for transcription, which is a common surface for indirect prompt injection if downstream systems or agents process the resulting transcripts without sanitization.
Ingestion points: The Lambda handler in lambda/handler.ts ingests data from S3 buckets, and the Cloud Run service in server.ts accepts audio URLs via HTTP POST requests.
Boundary markers: There are no specific delimiters or instructions to the agent to ignore embedded commands within the transcription output.
Capability inventory: The skill environment possesses capabilities for file system modification and container management (docker, kubectl), though these are not directly invoked by the data-processing handlers.
Sanitization: The implementation uses standard JSON parsing and URL decoding but does not include text-based sanitization of the generated transcripts.
[COMMAND_EXECUTION]: The skill includes shell scripts and configuration files designed to execute system-level commands for application deployment.
Evidence: The deploy.sh script and instructions in SKILL.md utilize docker build, kubectl apply, and gcloud run deploy to manage production environments. These are standard operations for the skill's stated purpose of facilitating software deployment.

deepgram-deploy-integration