The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill configures evaluators to process untrusted data from production AI traffic, such as $ai_generation events, traces, and sessions.
Ingestion points: Untrusted content enters the evaluation context through the evaluation_events global in Hog scripts or as input to LLM judges.
Capability inventory: Evaluators can execute logic and produce events; the skill uses posthog:execute-sql and posthog:llma-evaluation-create for configuration and verification.
Boundary markers: There are no explicit instructions for using delimiters to isolate untrusted data from evaluation instructions.
Sanitization: No specific sanitization or validation of the ingested generation data is mentioned.
[COMMAND_EXECUTION]: Uses a platform-specific SQL tool to query event counts for verification.
Evidence: The skill uses posthog:execute-sql to perform SELECT count() queries to validate the volume of events matching evaluation conditions.
[DYNAMIC_EXECUTION]: Involves the creation and testing of user-defined Hog source code.
Evidence: Uses posthog:llma-evaluation-test-hog and posthog:llma-evaluation-create to execute and persist evaluation logic provided as strings.

creating-online-evaluations