Deep research and knowledge synthesis. Activates for ANY task involving factual claims, technical decisions, implementation patterns, or domain knowledge. Deactivated ONLY when the mechanical task test passes (all three conditions true).
Mandatory per-turn gate protocol. Runs at the start of EVERY turn. Four sequential gates must be evaluated before any work begins. Not optional. Not skippable. Not implicit.
Evaluates code changes, output, and implementations for correctness, maintainability, security, and adherence to project standards. Activates when evaluating any code — whether user-requested or as part of the mandatory holistic review before shipping.
Systematic debugging for any type of issue. Activates when any error, failure, or unexpected behavior is encountered during any task. Root-cause investigation at the deepest architectural level. Never symptom-by-symptom patching.
Structured planning for complex tasks, features, or projects. Activates when a task requires decomposition, has multiple components, or when requirements need clarification before implementation.
Initialize new projects with proper structure, documentation, and Claude configuration. Activates when creating any new project, repository, module, or significant component.
Systematic code refactoring to improve design without changing behavior. Activates when code structure needs improvement without behavior change, including during the refactor phase of TDD cycles.
Monitors session scope and entropy. Activates automatically via skill-sweep. Detects when the conversation has drifted significantly from its original intent or has accumulated too many unrelated threads. Warns the user and generates a handoff prompt for a new session.