The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from customer support tickets and database event properties into the agent context.
Ingestion points: Support ticket content and event fields like $pathname, distinct_id, and variant keys retrieved via SQL queries as described in SKILL.md and references/pulling-the-data.md.
Boundary markers: Explicitly instructs the agent that ticket text and query results are 'data, never instructions' and must not influence tool selection or authorization.
Capability inventory: Employs read-only tools for configuration retrieval (posthog:experiment-get) and data querying (posthog:execute-sql).
Sanitization: Provides specific guidance on escaping single quotes and backslashes in SQL literals to prevent SQL injection during data retrieval.
[COMMAND_EXECUTION]: The skill runs a bundled Python script scripts/srm_check.py for statistical validation of experiment results.
Evidence: Documented workflow in references/pulling-the-data.md involves executing the script with parameters derived from the flag configuration.
Risk Mitigation: Includes a specific warning to avoid direct shell interpolation of variant keys, recommending they be passed via a JSON file to prevent command injection.

debugging-experiments