learn-from-claude

Four-phase gated coding workflow (Explore → Plan → Execute → Verify). Encodes the actual reasoning discipline of top-tier models, not the idealized version. Trigger: any task involving code creation, modification, or refactoring. Also invoked automatically by task-router.

Systematic debugging protocol: root-cause identification before any fix, hypothesis-driven investigation, and structured failure recovery. Prevents the most common failure mode of weaker models — applying random fixes and hoping something works. Trigger: bug, error, crash, broken, wrong output, not working, unexpected behavior. Also invoked automatically by task-router.

Structured extraction of implicit reasoning from top-tier model sessions into explicit, reusable rules for weaker models. Use when a weaker model fails a task, when you want to mine a successful session for transferable patterns, or when you notice a gap between what you did and what you could explain. Trigger: "distill", "extract pattern", "why did opus do that", "learn from", "what did I do differently".

Auto-activated at the start of every non-trivial conversation. Parses true intent before classifying, enforces the pre-flight checks that weaker models skip, and routes to the correct workflow. Trigger: always — no keyword required.