A harness for long-horizon Claude Code work: goals decomposed into loop-runnable phases, hooks that catch the bad session, verified iterations, and decisions that stay yours.
Use when about to claim a task is complete, fixed, implemented, or ready to merge/deploy/push — and when the user asks "fatto?", "finito?", "possiamo mergiare?", or "quanto rischio a pushare?". Use before ending any session or loop iteration that changed code.
Use on iteration 0 of any /goal — when a goal has just been set, when GOAL.md exists without PHASES.md, or when a goal's phase decomposition must be rebuilt after a re-scope ruling.
Use when ending a work session that changed project state, when the user says "facciamo il punto"/"a che punto siamo"/"riprendiamo", when starting work in a project that has a HANDOFF.md, or after a crash/context loss when prior session state must be reconstructed.
Use at the start of every /loop iteration and every /goal work cycle — when resuming autonomous work from a schedule wake-up, when a loop prompt says "continue the loop / continua il loop", or when about to schedule the next iteration of long-horizon work.
Use at the end of substantive coding, debugging, or exploration work — before the completion report — and whenever the user asks "c'è altro che dovrei sapere?" or requests a risk scan of recent changes.
Use when the user reports something broken, pastes an error/log/traceback/console output, says "non funziona"/"lo fa ancora"/"c'è un bug", or when a previous fix did not survive re-testing. Also use before editing code in response to any failure you have not yet reproduced.
Use when starting any feature or product slice expected to take more than a session — when the user describes something to build ("facciamo", "creiamo", "voglio una pagina/feature"), before implementation planning, or when a goal phase lacks a written spec.
Use before creative work whose requirements or design are not yet agreed — new features, new components, behavior changes with open questions. Skip when an approved spec or ruling already fixes the design. Explores user intent, requirements and design before implementation.