At the synthesis node of a GRD, we traverse the skill tree against the user's query, pull top-3 relevant nodes (depth 2), and inject that knowledge into the final prompt - only if relevance clears the threshold.
Cheap one-shot queries (price, trending) skip it entirely.