The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file contains a dynamic context execution command (!npm list @deepgram/sdk...) designed to run at load time to determine the current state of the project environment.
[INDIRECT_PROMPT_INJECTION]: The skill instructions and scripts process transcription results from external audio sources via the Deepgram API. This creates a vulnerability surface where malicious instructions embedded in audio could influence the agent during processing.
Ingestion points: Audio transcription results processed in SKILL.md (Step 4) and references/implementation.md (A/B Testing Models).
Boundary markers: No explicit delimiters or instructions are provided to the agent to disregard potential commands found within the transcription strings.
Capability inventory: The skill allows access to Bash, Write, Edit, and Grep tools, which could be misused if an injection is successful.
Sanitization: The processed transcription data is not sanitized or validated before being used in comparisons or logged to the console.
[EXTERNAL_DOWNLOADS]: The skill scripts include commands to download and install the @deepgram/sdk package and execute API requests to api.deepgram.com, which is a well-known service provider. These operations trace back to the vendor's expected infrastructure.

deepgram-upgrade-migration