Learn

This is a two-agent system that scans your codebase to extract conventions and patterns, then generates rule files in .claude/rules/ for Claude Code to follow going forward. The orchestrator gathers context and filters results while a separate analyst sub-agent does the forensic work. It presents top findings ranked by impact score and asks before writing anything. Smart move is running this when you join a new project so Claude learns the existing patterns instead of fighting them. One thing to know: it explicitly skips low-impact findings and deduplicates against existing rules, so if your project is already well-documented you might get zero results. That's actually the correct behavior, not a bug.