The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the $feature_flag_called event stream ($feature_flag and $feature_flag_response properties) which are supplied by external users and clients.
Ingestion points: Data is ingested from the events table via the execute-sql tool in the Get oriented, Ghost flags, and Response-distribution shift sections of SKILL.md.
Boundary markers: The skill contains explicit instructions in the 'Untrusted data' section to treat event-supplied strings as untrusted data, though it lacks formal delimiter-based markers for the LLM when interpolating these values into subsequent tool calls or reports.
Capability inventory: The agent has the capability to execute arbitrary SQL queries (execute-sql), write to a persistent scratchpad (scout-scratchpad-remember), and emit reports to an inbox (scout-emit-report, scout-edit-report) as defined in SKILL.md.
Sanitization: The instructions explicitly mandate truncating and sanitizing keys before use in the scratchpad and reporting them as untrusted snippets to prevent them from being interpreted as instructions.

signals-scout-feature-flags