The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio recordings and converts them into text transcripts. If these transcripts are subsequently passed to an LLM (via the optional --llm claude flag), there is a risk of indirect prompt injection where malicious instructions embedded in the audio could influence the LLM's behavior.
Ingestion points: Audio and video files are ingested and transcribed via the scripts/transcribe.py script.
Boundary markers: The script mitigates injection by wrapping the transcript in a JSON string and prepending a clear instruction to the LLM to treat the data as an untrusted summary target, not as executable instructions.
Capability inventory: The agent can write transcript files to the local disk and execute the Claude CLI for summarization.
Sanitization: Transcript content is HTML-escaped using html.escape before being saved to Markdown files to prevent script injection in Markdown viewers.
[DYNAMIC_EXECUTION]: The transcription script utilizes dynamic module loading for standard Python libraries.
Evidence: scripts/transcribe.py uses __import__("uuid") and __import__("html") at runtime to generate unique identifiers and escape transcript content. While these are standard libraries, the use of dynamic loading is noted.
[COMMAND_EXECUTION]: The skill executes external command-line tools to perform its primary functions.
Evidence: scripts/transcribe.py uses subprocess.run to invoke the claude CLI for summarization. The implementation is secure, using a list of arguments rather than a shell string, and includes hardening flags such as --bare, --tools '', and --no-session-persistence to restrict the CLI's environment during execution.

audio-transcriber