Превращение наблюдений сессий в правила и скиллы

learning

compound-learnings is a Claude Code skill that transforms ephemeral session observations into permanent, reusable artifacts — rules, skills, hooks, and agent updates. It reads files from `.claude/cache/learnings/`, builds a frequency table of patterns, consolidates similar formulations into their most general form, and detects meta-patterns: if more than 50% of observations cluster around one topic, it recommends creating a dedicated skill rather than multiple rules. A decision tree determines the artifact type for each pattern — step-by-step sequences become skills, event-triggered behaviors become hooks, and "when X do Y" heuristics become rules. Artifacts are only proposed when signal is strong enough (3+ sessions) and require explicit user approval via AskUserQuestion. Ideal for developers and teams who want to systematically compound knowledge from Claude sessions without losing recurring insights between contexts.