The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted meeting transcripts from various external formats including VTT, SRT, TXT, DOCX, and Google Docs. Malicious content within these transcripts could attempt to override the summarization instructions.
Ingestion points: Processes transcripts from local directories (REFLEX_TRANSCRIPT_SRC_DIR) and Google Workspace integration.
Boundary markers: The system prompt provided in the instructions does not use clear delimiters (like XML tags or triple quotes) or specific "ignore embedded instructions" directives to isolate untrusted data.
Capability inventory: The skill writes original and processed transcripts to the local file system (REFLEX_TRANSCRIPT_DST_DIR) and stores the final summary in a Qdrant vector database.
Sanitization: Preprocessing is focused on format cleanup (stripping timestamps and sequence numbers) but does not include sanitization of text for potentially malicious instructions.
[EXTERNAL_DOWNLOADS]: The skill relies on external libraries for specific file format processing.
Evidence: The documentation specifies extracting text from DOCX files using the python-docx library.

transcript-summarizer