Claude Projects let one coordinator fan work into parallel cloud threads

At a glance

Claude Corner tip

• Brief one coordinator with a single goal, then let it fan work into parallel Claude Code cloud threads.
• Put durable decisions in shared project memory so later threads inherit them instead of relearning them from you.
• If you do not have the redesigned Projects UI yet, wait or stay on one cloud session. Do not fake a coordinator with five Desktop chats.

Projects used to be a folder of files. As of 17 Sep 2026 they are a conversation: you brief one coordinator, it fans work into parallel Claude Code cloud threads, and those threads share memory while you close the laptop. That is not plan mode and it is not another skills trick. It is a different control surface. If you keep prompting every worker yourself, you are doing the coordinator’s job.

Why it matters

A redesigned Project is two layers. The coordinator chat scopes the goal, starts or reuses threads, reviews what comes back, and assembles the result. Each thread is a full Claude Code cloud session on its own branch and copy of the repo, so overlaps show up as ordinary merge conflicts. Threads can split further with subagents, loops, and workflows.

Shared memory is the part most people will underuse. Every thread writes into it and every new thread reads it, so “release moved to Friday” or “ask billing before touching invoices” survives the next fan-out. A project library collects files you add and artifacts threads produce.

Cost and access are the real constraints. Several full sessions at once will hit usage limits faster, and you can set model and effort separately for the coordinator and for workers. Threads run in the cloud today, so they cannot see files that live only on your machine or services behind a private network. Local threads are “coming very soon.”

Beta is gated. Select Claude Pro and Max subscribers who already use Claude Code cloud sessions and have no existing projects on web or desktop can start now at claude.com/projects. Everyone else on those plans should join the waitlist. Existing old-style projects keep working until the rollout reaches chat and Cowork. If you do not have the new UI yet, do not fake it with five Desktop chats. Wait, or use one cloud session until the coordinator exists.

How to brief one goal that naturally splits

1. Confirm `claude –version` is current (v2.1.277+ is fine). Open Claude Desktop or claude.com/projects. If you still see a static file folder, you are not on the redesign yet. 2. Start a project. Pick one goal and the repo (or repos) it needs. Configure the cloud environment, connectors, plugins, instructions, and model before the first fan-out. 3. Write the opening brief like you would to a chief of staff, not like a task list for five interns. Example: Goal: retire the v1 checkout endpoint across api, web, and mobile. Constraints: no billing-service edits without a human checkpoint. Open a PR per repo, run tests, then tell me merge order. 4. Let the coordinator propose work. Accept or cut threads. Do not spawn a second project for the same goal. 5. Stay in the project chat. Dive into a thread only to steer one conflict or to answer a check-in. Ask it to change check-in frequency or how often it starts new threads if the board gets noisy. 6. When PRs land, treat same-file overlap as a normal merge conflict. Do not tell two threads to “just rewrite HEAD.” 7. Open Project settings and read shared memory. Delete anything that is wrong. Wrong memory will be reused by every later thread. 8. Check project usage before you start another parallel goal. If you are near the cap, finish or park threads first.

Gotchas

• Cloud only for now. Internal APIs, localhost, and on-disk secrets are out of reach until local threads ship.
• One goal, many threads. Five unrelated goals in one project produces a shared memory junk drawer.
• Usage burns faster than a single Desktop session. Set worker effort down if the goal is mechanical. Keep high effort for the one thread that has to make a design call.
• If two threads edit the same module, you wanted a sequential brief, not a parallel one. Split by repo or by file ownership in the opening message.

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever you use (project chat if you have the redesign, otherwise a cloud session).

You are the coordinator for this repo, not a single worker. Goal: [one outcome, e.g. retire the v1 checkout endpoint across api/web/mobile]. Repos in scope: [list]. Hard constraints: [e.g. do not edit billing without asking me]. Definition of done: one PR per repo, tests green, a merge-order note in this chat. Work the way redesigned Claude Code Projects work: split into parallel threads only when work is isolated; give each thread its own branch and never two threads on the same files; write durable decisions into shared project memory and show me the memory notes before you spawn more threads; check in here instead of waiting for me to poll every worker; if you cannot reach local files or a private network, stop and say so. First reply: proposed thread list, files each thread may touch, and any constraint that blocks a thread. Wait for my yes before starting.