The Agent Skills Directory

[PROMPT_INJECTION]: The skill describes an architecture for extracting data from user conversations using an LLM. The example prompt template for extract_with_llm does not use explicit boundary markers to delimit untrusted conversation text from the instructions, creating a surface for indirect prompt injection where a user might attempt to manipulate CRM updates. However, the skill incorporates design patterns to mitigate this, including confidence-based filtering and human verification steps.\n
Ingestion points: conversation data processed in SKILL.md.\n
Boundary markers: Absent in the illustrative LLM prompt template.\n
Capability inventory: Includes logic for crm_client.update to modify CRM records.\n
Sanitization: Employs CONFIDENCE_THRESHOLDS and should_update validation logic to ensure data quality and prevent unintended overwrites.

custom-field-population