Claude’s Projects feature has been completely redesigned, and now it’s more like a project manager that hands out work on its own.
The old Projects was just a folder: you put in materials and wrote instructions, and chats could read them. The new version has launched in beta in Claude Code and become a long-running orchestrator conversation. You only state the goal; breaking down tasks, starting work, following up and reviewing results are all handed to Claude.
I went through the official docs, so here are the highlights.
The core change is an orchestrator + parallel threads. You toss requirements, bugs, errors and ideas into the project conversation, and it decides on its own which to answer directly, which get a new thread, and which go to a thread already working on something similar. Each thread is a full Claude Code cloud session with its own context, branch and copy of the repo. If two threads change the same code, it’s handled like a regular PR merge conflict.
All threads read and write the same project memory. Say “remember the release moved to Friday” or “ask me before touching the billing service”, and every thread opened after that starts work with it in mind. It also remembers your habits, like how often to report and how detailed updates should be.
You can also write project instructions, up to 16000 characters, which every new thread receives. Good for spelling out: what the goal is, which branch to pull from, which tests to run before finishing, what to do when it lacks permissions, and which actions it must ask you about first.
Alongside there’s an Overview panel with threads grouped by status: needs review, waiting on you, in progress, merging, idle, done. Step away for an hour, come back, and you can tell at a glance which PRs are ready for review and which thread is stuck. The Library holds the files you uploaded and the files Claude produced.
Once a thread opens a PR, it keeps watching it: if CI fails it pushes a fix itself, if there are review comments it makes the changes itself, and it notifies you once checks pass.
For tasks that need a local database, a simulator or your company’s intranet APIs, you can have just that thread run on your computer while everything else stays in the cloud as usual. This requires Claude Code v2.1.280 or later.
The orchestrator and threads can each have their own model and reasoning effort. Everything defaults to Opus, and in settings you can see how many tokens each thread has used.
Getting started is easy. You can get in from http://claude.ai/code, the Code tab in the desktop app, or the mobile app. Click Projects → New project in the left sidebar and fill in a name to create one; the goal and context are optional, and you can attach GitHub repos, files and Google Drive folders. Existing cloud sessions can also be upgraded directly by choosing Continue as project from the menu. If code is involved, the repo needs to be on http://github.com with the Claude GitHub App installed.
Once it’s set up, I’d suggest writing the project instructions first and throwing it a small task to test the waters: see how it reports and what it does on the branch, and only hand out work in bulk once you’re happy. These lines work really well:
- “List the threads you plan to open first, and wait for my confirmation before starting”
- “Run at most two threads at a time”
- “Only report when you’re done or stuck”
- “Use a small model for this one”
- “Tell me what the problems are first, don’t fix them yet”
It can do more than I expected:
- Retire an old endpoint: open one thread each for API, Web and mobile to migrate the callers, run tests, open PRs, and finally tell you which to merge first
- Set a performance target and have it profile endpoint by endpoint, try optimizations and open PRs
- Paste in bugs and errors for a service whenever they come up, and the pitfalls it hits get saved to memory
- Drop in a requirements doc and have it split into multiple threads that build the pieces separately
- It works without code too: upload a pile of contracts or customer support tickets, have it summarize the most common issues, and the results go straight into the Library as files
There are pitfalls, of course. The most obvious is how fast it burns through your quota: every thread is a full session, Claude decides how many to open, Pro users will really feel it, and idle threads watching PRs also get woken up by CI failures and keep consuming. If something gets written to memory wrong, every thread after that will get it wrong too, so you need to check it in settings regularly. Cloud threads can’t use the skills and MCPs installed on your machine; you have to commit them to the repo or configure them in project settings. For now projects are just for you and can’t be shared, and they aren’t in the terminal CLI or VS Code yet.
Right now it’s a public beta for Pro and Max, rolling out gradually, with priority for accounts that have used cloud sessions and have no old projects in chat or Cowork. Team and Enterprise aren’t available yet. If you don’t see the entry in the sidebar, you can join the waitlist, and the old Projects in chat still work as usual.
My take: it automates the whole “spin up a bunch of agents and play foreman yourself” thing. But the orchestrator’s ceiling depends on the brief you give it: the clearer you write the goal, the better it works.
