← ClaudeAtlas

orchestratelisted

Use when one or more implementation plans are ready and the user wants them built in separate builder sessions while this session stays the coordinator — types `/orchestrate [plan path ...] [--contract <path>] [--opus]`, or says "orchestrate ini", "kerjakan plan 1", "jalankan plan ini di session lain", "handoff ke builder", "session ini planner/parent, buat child-nya", "task cross repo", "BE dan FE sekaligus", "kerjakan di 2 repo", "jadi parent untuk ini". Especially use it for cross-repo work (backend + frontend, service + client) where several repos must agree on one API contract. Not for writing the plans (writing-plans) and not for interval loops (/loop).
zahidakhyar/skills · ★ 0 · AI & Automation · score 72
Install: claude install-skill zahidakhyar/skills
# Orchestrate This session is the **coordinator**. It hands each plan to its own child session (one per repo), keeps the shared contract, relays changes between children, and reviews the combined result. It never edits code itself: once it starts building, its context fills up and nobody is left holding the overall picture. One plan in one repo is just the N=1 case of the same flow. Children execute their plan **inline** (`superpowers:executing-plans`) by default, whatever execution skill the plan recommends. This skill already reviews at the end (step 5 review-fix loop, step 6 seam check, the coordinator's diff review, the child's advisor), so the per-task review of `superpowers:subagent-driven-development` mostly duplicates it, while costing a cold-start subagent per task and handing tasks it judges mechanical to a cheap model. Inline, the child's own model writes every line, so the model choice below is what actually shapes the code. Exception: a plan with more than 10 tasks uses `subagent-driven-development`, because one context carrying that many tasks degrades by the last ones, and a fresh subagent per task is worth its cost there. Children run on Sonnet at `high` effort. Quality comes from three backstops rather than a pricier default: the child's advisor (a stronger reviewer model, from `advisorModel` in the user's Claude Code settings; without it the checkpoints are skipped) at fixed checkpoints in the builder prompt, a review of each finished branch by a fresh r