This is Claude operating as a disciplined execution engine. You give it a pre-approved plan (typically in `docs/.plans/YYMMDD-HHmm-<plan-slug>/SUMMARY.md`), and it works through each phase exactly as written, no scope creep, no silent redesigns. It marks phases in progress, verifies each step, logs surprises and decisions directly in the plan file, and stops cold if something's ambiguous or blocked. Choose interactive mode for risky stuff like migrations or auth changes, batch mode for straightforward refactors. When done, it generates an execution report and asks if you want to verify, archive, or auto-commit. Built for multi-hour implementations where you need resumable checkpoints and audit trails, not ad-hoc fixes.
npx skills add https://github.com/buiducnhat/agent-skills --skill execute-plan