The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and ingest data from untrusted external sources, specifically 'ANY meeting recorder' transcripts (webhook payloads, share links, audio transcriptions, or manual pastes).
Ingestion points: Raw transcript data enters the agent context during the 'Normalize the input' phase and is processed through multiple verification phases.
Boundary markers: The instructions do not specify the use of delimiters (e.g., XML tags or triple quotes) to isolate the untrusted transcript content from the agent's instructions, increasing the risk that embedded commands could be followed.
Capability inventory: The skill has significant write capabilities, including put_page (modifying meetings/, people/, and companies/ directories), add_link, and add_timeline_entry.
Sanitization: While the skill mandates redacting PII and secrets (e.g., API keys, tokens) using a runPrivacyLint model, it does not explicitly filter for malicious instructions or prompt injection patterns within the transcript text.
Mitigation: The skill partially mitigates this risk through a rigorous 'Consistency check' and 'Verification checklist' (Phases 6 and 9) that require grounding every claim in verbatim transcript evidence and checking against existing known data.

meeting-ingestion