Cursor Multitask playbook¶
How to run Cursor Teamwork / Multitask on bounded features without polluting git with ephemeral orchestration state. Distilled from the aifeatures-site Vitest suite run (Aug 2026).
Canonical agent setup: Cursor agent setup.
When to use Multitask¶
Use Multitask when:
- The feature has clear milestones and acceptance criteria
- Parallel exploration or review gates save time (test suite, large refactor, multi-module feature)
- A single agent session would lose track of scope
Use a single Agent for doc fixes, one-file bugs, and config tweaks.
Use Ship workers when work must land as a Plane ticket + Forgejo PR + Woodpecker CI without your IDE open.
Dual-track pattern¶
Orchestrator (metadata only — no src/ edits)
├── Implementation track (per milestone)
│ Explorer → Worker → Reviewer → Challenger → Auditor
└── E2E / acceptance track
e2e_testing_orch → TEST_INFRA.md → TEST_READY.md
Role constraints¶
| Archetype | May do | Must not |
|---|---|---|
| Orchestrator | Update .agents/*/PROJECT.md, BRIEFING.md, DISPATCH.md, GATE_STATUS.md |
Edit src/, run builds/tests directly |
| Explorer | Read codebase, write survey notes | Implement features |
| Worker | Implement code + tests per milestone | Skip tests, hardcode fake assertions |
| Reviewer | Code review, request changes | Merge without green tests |
| Challenger | Adversarial cases, benchmark disputes | Override auditor veto |
| Auditor | Forensic integrity check, veto cheating | Write production code |
Starter template pack¶
Copy from templates/multitask/ into the target repo before starting Multitask:
| File | Purpose |
|---|---|
ORIGINAL_REQUEST.md |
Verbatim user prompt + acceptance criteria |
PROJECT.md |
Milestones, feature inventory, interface contracts |
BRIEFING.md |
Orchestrator mission + team roster stub |
DISPATCH.md |
Per-milestone dispatch log |
GATE_STATUS.md |
Pass/fail gates per milestone |
archetypes.md |
Role constraints (this table expanded) |
After the run:
- Commit distilled specs under
docs/(e.g.docs/testing/) - Commit source code and tests
- Do not commit
.agents/teamwork_*runtime trees — add.agents/to.gitignore
Example: aifeatures-site test suite (Aug 2026)¶
Prompt: Ultra-fast (<15s) Vitest + RTL suite in Docker for FeatureMatrix, CalculatorView, CommandPalette, ToolCard + glass-token styling.
| Milestone | Scope | Status |
|---|---|---|
| M1 | Test infra (vitest, happy-dom, renderWithRouter) | Done |
| M2 | Component + DOM assertion tests | Done |
| M3 | Translucency / glass token tests | Done |
| M4 | Docker benchmark <15s | Open (18–20s observed) |
| M5 | Tier 5 adversarial hardening | Planned |
Committed specs: aifeatures-site/docs/testing/ (from multitask TEST_*.md).
Multitask-ready work packages¶
Independent tracks for parallel subagents after docs land:
Documentation (low risk)¶
- Hub doc + per-repo
docs/AI_CONTEXT.mdIDE tables AGENTS.mdcross-links- Reviewer pass on MCP names and graphify commands
Cursor config bootstrap¶
.cursor/rules/+hooks.jsonfor app reposlocal/gen-*-mcp-json.shscriptstemplates/cursor-app/for onboard
Product (example: aifeatures M4)¶
- Rebuild Docker dev image
- Run
docker compose run --rm dev npm test -- --legacy-peer-deps - Document wall-clock in
docs/testing/benchmark.md - Challenger reviews <15s target vs measured time
Failure handling¶
Orchestrator escalation order (from aifeatures pattern):
- Retry same worker with clearer dispatch
- Replace worker
- Skip non-critical milestone (document in
GATE_STATUS.md) - Redistribute work across milestones
- Redesign scope in
PROJECT.md - Escalate to human (you)
Related¶
- Writing daily agent — portfolio scheduled content (Automations + Ship, not Multitask)
- Agent patterns to borrow — Shep, Fullsend patterns for ship UI
- Idea → Done — platform orchestrator vs IDE Multitask