Notes8 min read

Claude Code Projects: what they are and how to use them

A Claude Code project is one ongoing chat where Claude acts as coordinator. You describe work, and it starts a separate thread for each task, keeps shared notes, and shows what needs you. To start one, you need a Pro or Max plan and access to the beta at claude.ai/code.

Claude Code Projects are a new way to run several pieces of work at once without managing each one yourself. Anthropic announced them on September 17, 2026, as a beta. The short version: you talk to one long-running chat, and Claude decides what to hand off to helpers. You stay the director. It does the routing.

This note explains what a project is made of, when one is worth setting up, and how to keep it from eating your plan. The same ideas show up in directing a crew of agents, now built into a product.

How Claude Code Projects work

A project has two kinds of parts.

The project conversation. This is one chat that never really ends. Claude acts as the coordinator here. It reads what you send, decides what becomes a thread, and keeps track of every thread it started. It sees what threads report back, not every step they take.

Threads. These are the workers. Each thread is its own Claude Code session with its own context, and it does one piece of work and reports back. Most threads run in the cloud, which means they keep going after you close your laptop. A cloud thread that changes code works on its own branch, and opens a pull request when the job calls for one.

Anthropic's launch post sums it up as "threads do the work, Claude directs it." In their words, you brief the project the way you'd brief a chief of staff. You don't say which thread gets what. Claude picks.

A third piece ties it together: the Overview pane. It sits beside the chat and groups threads by state, such as Ready for review, Waiting on you, Working and Resolved. When you come back an hour or a day later, it's the first place to look.

Anthropic's short demo shows threads opening from one project conversation:

Anthropic's demo of project threads in Claude Code. Video: Anthropic.
One inbox, many threadsStandby
Requests land in one chat, and Claude answers small ones in place, starts a thread for new work, or hands a follow-up to a thread already on it.

Projects, sessions and the old Claude Projects

The name is confusing, so here's the split.

  • A normal Claude Code session is one conversation. You are the coordinator if you run several.
  • The older Claude Projects group chats, files and instructions. Each chat stays a separate conversation. Anthropic says those keep working as they do today until the redesign reaches them.
  • Claude Code Projects add a coordinator and threads on top. The coordinator works from recent messages, recent threads and project memory.

VentureBeat put the difference neatly: the older kind helps Claude remember the context you gave it, and the new kind helps it coordinate the work it's doing.

What a project remembers

Every cloud thread reads and adds to shared project memory. Anthropic's examples are small, human things: the release moved to Friday, why a feature was dropped, who to check with before touching billing.

You set two kinds of standing context once:

  • Project memory. Notes Claude keeps as files. You can say "remember this" in the chat, and you can read, edit or delete the notes in project settings.
  • Project instructions. The brief every new thread starts from, up to 16,000 characters. A good one covers what the project is for, where the work happens, how a thread checks its own work, what to do when something is missing, and what needs your go-ahead first.

Here is a short brief you could adapt. The last two lines are the ones that save you the most:

This project turns customer feedback for Kudos into weekly write-ups.
- Read only the files in the Library tab.
- Before you call work done, list every quote you used and the file it came from.
- If a file you need is missing, say exactly which one and stop. Don't guess.
- Don't send or publish anything. Save drafts as files and wait for me.

This is the same habit as a skill: write down once what you'd otherwise explain every time.

When a project is worth it

The docs give a clear test. A project fits when the work has a goal that outlasts one session and keeps producing tasks. Their examples:

  • One goal across several code repositories
  • An area you keep feeding, such as bug reports for one service
  • A build or move bigger than one session
  • Work that isn't code, such as a folder of contracts or a support-ticket export you keep asking new questions about

That last one matters if you don't write code. A project doesn't need a repository. You can upload files, and threads can research, write documents and deliver them as files on the project's Library tab.

Skip a project when one task fits in one session, when every task needs something only your computer has, or when a single job just repeats on a schedule. For that last case, the docs point you to a routine.

Try it: your first project

You'll need a Pro or Max plan. Projects are in public beta there, rolling out gradually, and they aren't available on Team or Enterprise plans yet. Open claude.ai/code and look for Projects in the sidebar. If it isn't there yet, the docs say you can join a waitlist.

  1. Pick one stream of work. Choose something you'll keep adding to. For a made-up testimonial app called Kudos, that could be "everything customers say about us."
  2. Create the project. Select Projects, then New project. Only the name is required. A goal and context are optional, and you can add them later.
  3. Write the instructions first. Open project settings, go to Memory, and paste a brief like the one above.
  4. Send one small task. Don't send the big batch yet. Open the thread when it finishes and see how it reported back. Did it guess anything?
  5. Fix the guess in memory. When you correct a thread, also tell Claude to remember the correction, so later threads start with it.
  6. Then send a batch. Ask Claude to propose threads before starting them, and to run only a few at a time. Drop those limits once the results match what you'd do yourself.

Keep your plan from draining

Several threads at once means several full sessions at once. Anthropic warns that projects can reach your usage limits faster, and the docs say that on a Pro plan you should expect to hit them sooner on days you run one. If you haven't read it yet, what Claude Code really costs explains how limits work.

What the docs say you can do:

  • Check the defaults. A new project runs Opus everywhere, at high effort for threads and low effort for the conversation. For work that doesn't need the biggest model, choose a smaller model or lower effort in project settings.
  • Open Usage in project settings. It shows token use by thread and model, and how much the coordinator itself used.
  • Start fresh threads. A follow-up sent to a thread that has been idle for over an hour makes it re-read its whole conversation first. For new work, a fresh thread can cost less than reviving a big old one.
  • Stop PR watchers. An idle thread that is watching a pull request wakes up when tests fail or a comment arrives, and that uses your plan. Ask it to stop watching if you don't want that.
  • Ask for fewer threads. There is no fixed number of threads. The one enforced limit is 200 new threads per day across your projects.

Good news: a project with nothing running and no new messages doesn't use your plan while it sits idle. And if a thread hits your plan's limit, it waits and continues on its own when the limit resets. It can't spend past your limits unless you've turned on usage credits.

What to know before you lean on it

This is a beta, and the docs list real limits:

  • Projects live at claude.ai/code, in the Claude desktop app and in the mobile app. They aren't in the terminal, the VS Code extension or JetBrains.
  • If threads touch code, the code needs to be on github.com, with the Claude GitHub App installed on it.
  • Cloud threads don't pick up the skills and tools installed on your own computer. To give them a skill, commit it to a repository in the project.
  • If two threads change the same code, it's a normal merge conflict, handled like any other pull request.
  • A project belongs to one person. You can't share it yet.
  • Running a thread on your own computer is possible through Remote Control, but only while that computer is awake.

The review step doesn't go away either. A coordinator can route work and a thread can open a pull request, but you still decide what's good. That part of the job, the taste and the go or no-go, stays yours. In Chief of Agents, that's what we practice: directing a crew of agents toward a launched product with paying customers, and judging what comes back.

Sources

Learn this inside Claude Code.

Join the waitlist for Chief of Agents. You'll get the free beginner's handbook and skills right away, and hear first when the course opens.

One email to confirm, then the handbook and skills. Unsubscribe anytime.

Keep reading

  1. 6 min read

    Claude Code for designers: what it is, and why it isn't just for engineers

    What a coding agent does, why directing one is a design skill, and a safe first hour for designers new to Claude Code.

  2. 7 min read

    Agents, skills, MCPs and hooks, explained with Figma

    Agent vs skill vs MCP, plus subagents, hooks and plugins: seven Claude Code terms, each matched to something you know in Figma.

  3. 7 min read

    Connecting Claude to your tools safely: MCP for non-developers

    Is it safe to connect Claude to Gmail or Notion? What an MCP server can see and do, why to start read-only, and which hooks help.

All notes →

© 2026 Chief of Agents · Notes · PrivacyIdeate. Build. Sell. You lead, the agents work.