Claude's New Projects Run the Work. Checking It Is Still Your Job.
On the morning of 5 October a dialog box opened over my Claude Projects page. “Before you continue, upgrade your existing projects or archive the ones you don’t need.” I had 30 projects, between them holding 93 files, 11 sets of standing instructions and 24 project memories. The box offered two buttons, “Not now” and “Upgrade 30 projects”. It didn’t say what the upgrade would change. I clicked the second button a little after 11:15.
At 12:11 I opened Zidi-UG, the project for a financial dashboard I’m building. Claude greeted me with “I moved over everything from your project, but things look a little different.” A card listed what had come across. One line read “13 chats”, with “chats” struck through and “threads” written beside it.
Projects in this new form are in public beta on Claude’s Pro and Max plans and are still rolling out, so you may not see them yet. They aren’t offered on Team or Enterprise plans (as of 6 November 2026, Projects documentation).
That strikethrough marks a real change in how the tool works. An old project was a folder of separate chats that shared some files. A new project is one ongoing conversation in which Claude takes the work I send, splits it into jobs, and hands each job to a separate working session. Most of the coordinating has moved from me to Claude. The checking hasn’t, and for anyone who signs off on work, that’s the part that matters.
How I got here
| Year | What the tool did | Milestone |
|---|---|---|
| 2021 | Finished your line | GitHub Copilot technical preview, 29 June |
| 2022 | Answered you | ChatGPT, 30 November |
| 2023 to 2024 | Remembered the conversation | OpenAI Assistants API threads, 6 November 2023; Claude Projects, 25 June 2024 |
| 2025 | Did the work | Claude Code: preview 24 February, general release 22 May |
| 2026 | Ran other sessions | Cowork, 12 January; agent teams, 5 February; routines, 14 April; redesigned Projects, 17 September |
My own version starts in VS Code around 2021, with an extension that suggested the rest of a line as I typed. I can’t remember what it was called, which says a lot about how those tools felt. I used ChatGPT on the day it came out and haven’t gone a day without an AI tool since. For most of that time I was the coordinator. Anthropic’s documentation describes the old routine exactly: “you decide what each one works on, repeat the same background at the start of each, and check back to see which finished or needs an answer” (Projects documentation). The redesign hands that routine to Claude.
What a project is now
Think of the project conversation as a resident engineer. You give the resident engineer a brief. They split it into jobs and send a separate crew to do each one. Each crew keeps its own notebook, finishes its job and reports back. The resident engineer keeps the overall picture and tells you when something needs your decision.
The documentation calls the conversation the coordinator. It “takes what you send, decides what becomes a thread, and keeps track of every thread it started.” A thread is “a separate session with its own context window that does one piece of work and reports back.” The context window is how much a session can hold in mind at once. A thread usually runs on Anthropic’s computers, so it keeps going after I close my laptop.
This article was written that way. I uploaded my research notes to the Mwangaza project and asked for the piece. The coordinator started one thread to write it. When I asked for a usage test, it started a second thread, which wrote its results to a file the first thread then read. I never set either one up by hand.
The analogy breaks in one useful place. Crews on a real site talk to each other. Threads mostly don’t. The coordinator “sees what threads report back, not every step they take,” and the threads don’t see each other’s notebooks. That’s also the real limit. When Anthropic’s engineers built a system of cooperating agents, they found it a poor fit for work where agents must share context or depend heavily on each other, “such as most coding tasks” (Anthropic).
Five other parts of a project are worth knowing.
The Overview pane sorts threads by state: Ready for review, Waiting on you, Working, Landing, Idle and Resolved. Waiting on you is the one to check. It holds threads that need a reply or an approval, or that failed. A thread is marked Resolved by me, by Claude once I’ve taken the last step, or automatically after a week with no activity, and I can reopen it.
Project instructions are standing rules sent to every new thread, up to 16,000 characters. A rule written once reaches every crew.
Project memory is a set of files Claude keeps about the project: decisions, requirements, pitfalls. Every thread reads the index file when it starts. The coordinator works from recent messages, recent threads and memory, not the whole history, so anything that must never be forgotten belongs in memory.
The Library holds the files I add and the files threads produce. A project doesn’t need any code. The documentation names “work that isn’t code” as a use, such as a folder of documents you keep asking new questions about, with each answer delivered as a file.
Repositories are optional. A repository is a project’s set of files kept on GitHub. A thread working in one makes its edits on a separate working copy, called a branch, and offers them back as a proposed change waiting for approval, called a pull request. Nothing reaches the main copy until someone approves it.
What happened to my 30 projects
Before clicking upgrade I copied down what the dialog said each project held. That gave me a baseline.
In Zidi-UG the counts matched. Thirteen old chats became thirteen threads, twelve marked Resolved and one Idle. The instructions had moved to the project settings. The old project memory had become five memory files.
The Mwangaza project came through the same way. The dialog had listed 4 files and project memory. The upgraded project has the same 4 files, 5 memory files, and 30 threads: 27 from old chats and 3 from old Cowork tasks. I opened a chat from August and the full text was there, links included. The original project still exists in read-only form, so the upgrade copied my work rather than moving it.
Here’s what I haven’t checked. I’ve looked closely at 2 of 30 projects. I found no documentation on reversing the upgrade, and a read-only original isn’t the same as an undo button. The progress box wasn’t reliable either. It showed 2 of 30 done across three checks, the last at 11:54. At 12:11, Zidi-UG, still listed there as “In progress”, was open and working.
My first reaction, written down at the time, was “We are not keeping up with the changes.” Two projects in, I’d still say that.
What it costs, and when not to use it
Every running thread is a full working session, several can run at once, and the coordinator uses its share too. The documentation says a project “draws on the same plan limits as your other Claude Code sessions and uses them faster.” The cloud session credit Anthropic gave Pro and Max subscribers this autumn ($100 on Pro, $250 on Max, expired 4 November) didn’t help: its terms list “Projects and Routines” under what the credit doesn’t cover (Anthropic Help Center). My own usage screen agreed. Between 13:30 and 15:07 my session bar climbed from 15% to 29% while the credit stayed at $250. A project that sits idle, with nothing running and nothing watched, uses nothing.
The documentation is also plain about when a project is the wrong tool. One task that fits in a single session should stay a session. Work where every task needs your own computer belongs on your computer. Most of the questions I ask in a day still belong in a plain chat.
The part that didn’t move
I haven’t read code line by line in almost six months, and I haven’t written a function from scratch in that time either. That sentence should worry anyone who has signed a calculation. The answer isn’t to go back to reading every line. It is to decide, before the work starts, what evidence you need to see before you sign.
The documentation’s own sample instructions point the same way. A thread runs the tests and pastes the summary lines, and one sample line tells Claude not to merge a change without asking. The resident engineer can run the crews now. Signing off on their work is still my job.
Sources
- Let Claude coordinate ongoing work with Projects, Claude Code documentation, Anthropic. Checked 6 November 2026.
- Cloud sessions bonus credit promotion, Anthropic Help Center. Checked 6 November 2026.
- How we built our multi-agent research system, Anthropic Engineering, June 2025.
- Introducing GitHub Copilot, GitHub, 29 June 2021.
- Introducing ChatGPT, OpenAI, 30 November 2022.
- New models and developer products announced at DevDay, OpenAI, 6 November 2023.
- Collaborate with Claude on Projects, Anthropic, 25 June 2024.
- Claude 3.7 Sonnet and Claude Code, Anthropic, 24 February 2025.
- Introducing Claude 4, Anthropic, 22 May 2025.
- Anthropic launches Cowork, VentureBeat, 12 January 2026.
- Claude Opus 4.6, Anthropic, 5 February 2026.
- Introducing routines in Claude Code, Claude blog, 14 April 2026.
- Projects redesigned: from folder to conversation, Claude blog, 17 September 2026.
How this was written: Drafted with AI (Claude), edited and verified by the author.