cursor-orchestratelisted

# Orchestrate a Cursor fleet (Claude plans, Cursor executes) Claude stays the brain: it owns architecture, UX/UI and design decisions, the plan, task decomposition, review/acceptance, and final integration. `cursor-agent` workers do the execution on the Cursor subscription's quota. This is agent calling at scale, not a model-backend swap. ## Step 1 — pick the mode Check the user's words first, then fall back to auto-detect by size: | Signal | Mode | |--------|------| | user said **"sub"** / **"subagent"** | **A-simple**: one `cursor-worker` subagent or a single `cursor_run` call | | user said **"workflow"** / **"w"** | **B**: the JS harness (`orchestrator.js`) | | **≤ 1** independent slice | A-simple | | **2–8** independent slices | **A-fan-out**: several `cursor_run` calls in parallel | | **> 8** slices, or multi-phase build, or needs retries / worktrees / determinism | **B**: the JS harness | State which mode you chose and why in one line. ## Step 2 — decompose (Claude's job) Cut the work into **independent vertical slices**. Each slice becomes one worker task. Write every task **fully self-contained** — the worker starts with a blank context and cannot see this conversation: > **File paths** to touch · **the goal** · **acceptance criteria** (how to know it's done) · **what NOT to change**. Reference the project's conventions by pointing at its `CLAUDE.md`/`AGENTS.md` (see Context below), not by re-explaining them. Bad: "add the auth we discussed." Good: "In `src/a