The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from collections, dashboards, and knowledge bases which may contain content from external users or automations, creating a surface for indirect prompt injection.
Ingestion points: The skill ingests potentially untrusted data via knowledge.search (documented in references/knowledge.md), collection.query_rows (documented in references/collections.md), and widget data in dashboard.get_page (documented in references/dashboards.md).
Boundary markers: The skill contains explicit safety instructions in SKILL.md (Rule 10) and references/knowledge.md directing the agent to treat search hits and returned text as reference data only, never as instructions. It also utilizes a gate mechanism to evaluate the confidence and trust level of search results.
Capability inventory: The skill possesses significant write and management capabilities, including elaichi__collection__upsert_rows, elaichi__dashboard__draft.apply, and various schema modification tools (elaichi__collection__update_schema).
Sanitization: The skill employs a human-in-the-loop review system. Data added via chat or automations is marked as "unreviewed" and requires a person to manually verify it in the console before it is considered trusted.

elaichi-data