Create a detailed implementation plan with test specs, phased task breakdown, and Agent Context blocks for agent-driven execution. Consumes a research artifact or works from a feature description directly. Creates a task graph in yaks, beads, or native tasks. Use when the user wants to plan a feature, create an implementation roadmap, break down work into tasks, or prepare for TDD implementation. Do NOT use for codebase exploration (use research instead) or executing a plan (use implement instead). Triggers on phrases like "plan the implementation", "create a roadmap", "break this into tasks", "plan before coding", "draft a plan", "turn research into a plan", or "create task breakdown".
Execute an approved implementation plan using task-graph-driven orchestration with isolated TDD agents (RED/GREEN/VALIDATE gates). Reads a task graph from yaks, beads, or native tasks and dispatches agents phase by phase. Use when the user has an approved plan and wants to start building, execute a task graph, resume an interrupted implementation, or pick up where a previous session left off. Do NOT use for planning (use plan-tasks instead) or codebase exploration (use research instead). Triggers on phrases like "implement the plan", "execute the plan", "start building", "resume implementation", "pick up where we left off", "dispatch agents", or "build from plan".
Explore a codebase and research external patterns before building a non-trivial feature. Spawns parallel subagents for codebase exploration AND web/pattern research, then synthesizes findings into a compact research artifact at .light/sessions/. Use when the user wants to understand existing code, investigate architecture, find best practices, or scope out a feature before planning or implementing. Do NOT use for debugging failures, fixing bugs, or post-implementation reflection. Triggers on phrases like "research the codebase", "explore before building", "understand how X works", "what exists for Y", "help me scope this feature", "deep dive into", or "before I start building".
Assess and strengthen the agentic coding environment surrounding a codebase. Audits instruction files, hooks, type safety, linting, pre-commit, architecture tests, CI gates, and sandbox permissions — then produces a gap analysis with actionable improvements. Use when the user wants to improve their Claude Code setup, harden their development environment for AI agents, add hooks or guardrails, assess harness maturity, or mentions "harness engineering". Also use when the user says things like "make my repo more agent-friendly", "add guardrails", "improve my Claude setup", "audit my dev environment", or "what controls am I missing".
Guides writing minimal Architecture Decision Records (ADRs). Use when recording architectural decisions, documenting design choices, capturing technical decisions with context and alternatives, or when user mentions ADR, architecture decision, or decision record.
Post-session skill to reflect on what was built and produce improvement proposals. Reads recent git history, artifacts, and context files to extract learnings for skills, CLAUDE.md, hooks, and plan templates. Use after any substantive session as a post-mortem or retrospective.
Boundary-focused TDD workflow enforcing L3/L4 altitude testing, property-based tests, and red-green-refactor phase separation. Use when starting a new module or feature with test-first discipline.