The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes AI-generated cluster titles and descriptions that are stored within the PostHog events database. These strings are ingested via the $ai_clusters property in SKILL.md and are used for analysis and display. While this represents a potential attack surface for indirect prompt injection from processed LLM traffic, the skill is designed for human-in-the-loop investigation and primarily uses the data for summarization and visualization.
Ingestion points: Analyzes AI-generated metadata from $ai_trace_clusters, $ai_generation_clusters, and $ai_evaluation_clusters events.
Boundary markers: Not explicitly defined for property values, though SQL query results are handled through structured tools.
Capability inventory: SQL execution (posthog:execute-sql), trace inspection (posthog:query-llm-trace), and local result processing via scripts/print_clusters.py.
Sanitization: The print_clusters.py script performs standard JSON parsing and string truncation for display purposes.
[SAFE]: All identified tools, domains (posthog.com), and resources originate from the verified vendor. The included Python script scripts/print_clusters.py is a utility for parsing local result files and contains no network or dangerous system calls.

exploring-llm-clusters