The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) because it processes untrusted meeting transcripts and interpolates them into LLM prompts.\n
Ingestion points: Transcript content is ingested via scripts/parse-transcript.sh from user-supplied files or standard input.\n
Boundary markers: The prompt in scripts/extract-items.sh uses a "TRANSCRIPT:" header to delineate input, but lacks strong delimiters or explicit instructions for the model to ignore potential instructions embedded within the transcript text.\n
Capability inventory: The skill utilizes local command execution, file system access for history logging, and network access to communicate with LLM APIs.\n
Sanitization: Transcript content is processed as raw text without specific sanitization for adversarial instructions before being sent to the LLM.\n- [DATA_EXFILTRATION]: Meeting transcripts are processed by transmitting them to external LLM providers.\n
Evidence: scripts/extract-items.sh and scripts/generate-outputs.sh send transcript content to official Anthropic or OpenAI API endpoints to extract action items and generate reports.\n- [COMMAND_EXECUTION]: The skill utilizes local bash scripts and inline Python code for data transformation.\n
Evidence: The orchestrator scripts/meeting-autopilot.sh and supporting scripts like scripts/parse-transcript.sh execute shell commands and Python logic.\n
Analysis: The implementation employs safe practices such as using jq for argument handling and reading from standard input in Python scripts, which minimizes traditional command injection risks.

meeting-autopilot