The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted user-generated content from open-text survey responses, creating a surface for indirect prompt injection attacks.
Ingestion points: Untrusted data enters the agent context via execute-sql queries that retrieve survey response properties (e.g., $survey_response_<question_id>) from the events table, as specified in SKILL.md and references/response-querying.md.
Boundary markers: The instructions lack explicit delimiters or markers to isolate the untrusted survey content from the agent's instructional context during the summarization and theme aggregation process.
Capability inventory: The agent has the capability to write to the project's reporting channel using emit_report and edit_report, and can store state via signals-scout-scratchpad-remember.
Sanitization: The skill provides mitigation strategies, including SQL-level filters to exclude short or common test responses (e.g., length(response) > 5 and NOT IN ('test', 'qwe', 'asdf')), and instructions for the agent to summarize themes rather than quoting raw text verbatim, especially when PII is suspected.

signals-scout-surveys