The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from production AI traces, which may contain adversarial content.
Ingestion points: The posthog:query-llm-trace tool reads raw content from $ai_input and $ai_output_choices (as described in SKILL.md Step 3 and references/finding-traces.md).
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions when reading trace content, potentially allowing the agent to be influenced by instructions embedded in the data.
Capability inventory: The skill utilizes posthog:execute-sql for data aggregation and posthog:query-llm-trace for detailed record inspection.
Sanitization: No sanitization or validation of the trace content is mentioned before it is presented to the agent for categorization.
[COMMAND_EXECUTION]: The skill leverages the posthog:execute-sql tool to run structured queries against the PostHog database. While restricted to the database environment, this allows the agent to execute arbitrary data-plane logic as defined in the SQL templates within references/finding-traces.md.

exploring-ai-failures