Cloud agent playbook

---
name: cloud-agent-playbook
description: Write a focused brief for a Cursor cloud agent: outcome, constraints, stack, and how to verify. Use when launching or rewriting an agent task.
---

# Cloud agent playbook

## Brief

State the outcome as a user-visible slice, not a platform:

- Who it is for
- What they can do when it is done
- What “done” looks like in the browser or CLI

## Constraints worth writing down

- Stack (or “use the repo default”)
- What not to add (auth, database, extra services)
- Domain or product name
- Independent / trademark notes if the work sits near another product

## Verification

Ask for:

1. A running surface (dev server or command)
2. Empty / loading / error where they apply
3. A README that matches reality

## After the brief (2026 harness)

- `/goal` for an objective that is not one edit (“CI green,” “empty states on every new list”).
- Custom Mode if the brief is really a playbook — pin the skill for the whole session.
- Subscriptions if work should wake on a PR, Slack thread, or schedule. Cloud agents already subscribe to PRs they open.
- Isolated subagent VMs when you want a clean checkout to test or swarm. Do not share one dirty working tree across colliding fixes.
- **Start from scratch** if there is no GitHub/GitLab repo yet. Prompt first, then Create repo on Origin when the slice is usable. Preview is port-forwarded; Vercel publish is optional.
- **Projects** when the work should keep memory for weeks. The coordinator plans and delegates; it does not write the feature.
- **Self-hosted** only when code, builds, or secrets must stay on your network. Cursor-hosted remains the default.

## Anti-patterns

- “Build a SaaS for X” with no first slice
- Asking for auth + payments + AI + admin in one run
- Leaving the agent to invent the product name and voice
- A daily automation with no definition of done