Claude Code Projects turn Claude into a coordinator. A project is one long conversation where you drop tasks as they come up; Claude starts a thread for each one, runs the threads in parallel as cloud sessions, and gives every thread the same instructions, memory and repositories. We have run this site's SEO and AI-answer work in a project since October 16, 2026. This guide covers how the beta works, checked against Anthropic's docs on November 16, 2026, and what a month of daily use taught us.
What are Claude Code Projects?
A Claude Code project is one ongoing conversation where Claude acts as coordinator for a stream of related work. You send it tasks, and it starts a thread for each one. Each thread is a separate Claude Code session, usually in the cloud, that does one piece of work and reports back. Every thread starts from the same instructions, memory and repositories.
Anthropic launched the redesign on September 17, 2026 with one line that sums it up: "Projects have threads that do the work and a coordinator that directs them." The old Projects in claude.ai chat grouped conversations and reference files. The new ones run work.
How does a Claude Code project work?
Four parts do the work, per Anthropic's Projects docs:
- The project conversation. Claude as coordinator. It decides what becomes a thread, answers quick questions in place, and sees what threads report back rather than every step they take.
- Threads. The workers. Each is a separate session with its own context window. A cloud thread works on its own branch and opens a pull request when the work calls for one, then watches it and pushes fixes when CI fails.
- Standing context. Every cloud thread starts with the project's repositories and files, its instructions and memory, the CLAUDE.md in each repository (not the repository's skills, .claude/settings.json or .mcp.json), the connectors on your claude.ai account, and a cloud environment.
- The Overview pane. Threads grouped as Ready for review, Waiting on you, Working, Landing, Idle and Resolved, plus tabs for the file Library, pull requests and scheduled tasks.
A new project runs Opus everywhere, at high effort for threads and low effort for the coordinator. You change both in Project settings. Threads run in auto mode, so most tool calls go through without asking you.
Who can use Claude Code Projects?
Projects are in public beta on the Pro and Max plans and rolling out gradually, according to Anthropic's docs as of November 16, 2026. They are not offered on Team or Enterprise plans yet. You use them at claude.ai/code, in the Code tab of the desktop app, or in the Claude mobile app, not in the terminal CLI or IDE extensions.
If Projects is missing from your sidebar, check under More. If it is not there either, the rollout has not reached your account and Anthropic offers a waitlist. Code work also needs your repository on github.com with the Claude GitHub App installed on it.
What we ran in one project for a month
Most guides to Projects are written from the announcement. This one comes from use. Our project's goal is plain: own "second brain" search for viableedge.com and turn it into course sign-ups. It has one repository (this site), about 11,000 characters of instructions, and a handful of scheduled tasks. Between October 16 and November 16, 2026:
| What | Count | Where the number comes from |
|---|---|---|
| Threads started | About 40 | The project's thread list |
| Pull requests merged | 60 | GitHub, PRs titled "SEO:" merged in the window |
| Articles shipped or rewritten | 34 | Our article log |
| Memory files Claude wrote | 32, plus the index | The project's memory folder |
| Google AI impressions per week | 783 to 1,971 | Search Console Generative AI report, weeks of Oct 13 and Oct 27 |
The work splits the way the docs describe. Scheduled tasks start threads on their own: a weekday 6:57am content run (new posts Monday, Wednesday and Friday, a rewrite of a page close to page one on Tuesday and Thursday), a Monday performance report, a Friday check of which AI engines cite us, and a Friday retrospective. Everything else arrives as a message in the conversation: an article worth assessing, a keyword idea, an audit export. Each becomes a thread with its own pull request, and the coordinator posts one line when a thread finishes or needs a decision.
How should you set up your first project?
Write the instructions before you send work. Ours started at a page and grew every week. The brief that matters covers five things, in this order:
- The goal, in one line. Ours names the search term and the business outcome, so every thread can tell useful work from busywork.
- Where work ships. Branch, pull request, the checks to run, and what "done" means. For us, done is a live production deploy, not a merge.
- What needs you. We delegated merges once CI is green and kept goal changes and spending past a budget for decision cards.
- Where data comes from. Connectors first, then your own computer. Our keyword tool runs on a Mac, so threads reach it through a connector, and fall back to a thread on that Mac through Remote Control.
- How to report. Short, outcome first, no closing questions.
Then send one small real task and read the thread end to end before sending a batch. Anthropic's docs suggest the same, and it is where we caught our first wrong assumption.
What is the difference between project memory and project instructions?
Project instructions are a brief you write, up to 16,000 characters, sent to every new thread and to the coordinator. Project memory is a set of files Claude writes itself as it works: decisions, requirements and pitfalls, indexed in a MEMORY.md file every cloud thread reads at the start. Put rules in instructions and learned facts in memory.
Repository rules, like build commands and copy rules, stay in the repository's CLAUDE.md, which every cloud thread loads. Our split: instructions hold how we work, memory holds what happened (budgets, rulings, which data source broke last week), CLAUDE.md holds how the site is built. When a correction should outlast one thread, ask Claude to remember it, and it lands in memory for the next thread.
What a month taught us
- "Blocked" has to mean "needs you". Threads waiting on CI or data showed as blocked and read like a to-do list for the owner. The rule now: a thread waiting on a machine resolves itself and wakes when the result arrives. Only decisions and approvals reach a person.
- Never end a turn on "waiting". A green pull request sat unmerged for 2.5 hours on November 10 because the thread stopped without anything set to wake it. Threads now keep a timer running while CI runs.
- Diagnose a missed run from its own error. On November 8 the morning post never shipped. We first blamed a late schedule. The run's own session said the subscription payment was past due. Read the failure before naming a cause.
- Let a retrospective edit the rules. Every Friday a scheduled thread reviews the week and changes the instructions directly when the change changes no goal and adds no new ask of the owner. That loop did more for quality than any single prompt.
- Watch the plan limits. Several threads at Opus high effort use a plan fast. Anthropic's docs say so plainly: on Pro especially, expect to reach your limit sooner on days you run a project.
Claude Code Projects vs cloud sessions, scheduled tasks and agent teams
| Feature | What it is | Use it when |
|---|---|---|
| Project | A coordinator that starts and tracks threads with shared instructions and memory | A goal that outlasts one session and keeps producing tasks |
| Cloud session | One Claude Code session in the cloud | One task that fits in a session |
| Scheduled task | A prompt that runs on a schedule | One recurring job with no conversation around it (inside a project, its runs become threads) |
| Agent team | One session that starts teammate sessions for a single task | A big task that ends when it is done |
| Remote Control | Connects claude.ai to a session on your own machine | A thread needs your local files, tools or network |
Our guide to AI agent teams covers the single-task version, and Claude Cowork vs Claude Code covers where Cowork fits.
Do Claude Code Projects cost extra?
No separate price. A project draws on the same plan limits as your other Claude Code sessions, and uses them faster because several threads run at once and the coordinator spends tokens of its own. A project with no running threads, no watched pull requests and no new messages uses nothing while it sits idle.
Project settings has a Usage page that splits token use by thread and by model. Anthropic enforces a cap of 200 new threads a day across your projects. To spend less, run threads on a smaller model or lower effort, and start a fresh thread for new work instead of reviving a large idle one.
Is a Claude Code project a second brain?
For one stream of work, close to it. Project memory, instructions and the file library are a knowledge base that Claude reads before every task and updates as it learns. The difference is scope: project memory knows the project, while a second brain holds what you know across all of your work. The two work best together.
That is the pattern we teach: a second brain kept as plain files an AI can read, the same way a company keeps an internal knowledge base. A project adds the part most second brains lack, an agent that keeps the knowledge current while it works. Our guide to keeping a knowledge base up to date covers that habit outside of Projects, and the Claude Code + Obsidian setup is the personal version.
To give a project your business in one afternoon, the Business Brain Starter Kit is six short files (your business, customers, voice, offers, one repeat task and your decisions) with a setup guide for the new Projects and one for Classic Projects. It is a one-time purchase.
What are the limits of the beta?
- A project belongs to one user. You can't share it or its threads.
- A thread can't move between projects, and two projects can't merge.
- The coordinator conversation has no connectors of its own, so work that needs one goes to a thread.
- A thread on your own computer starts with the instructions but not the memory files, and pauses while the computer sleeps.
- A cloud sandbox can be lost between turns, so long tasks should commit and push as they go.
Features change fast in a beta. We recheck Anthropic's docs and update this page when they do.
Where to start
Pick one stream of work you feed every week, write the five-part brief above, and send one task. If you want to turn your Claude Code setup into a service you sell, the Claude Code Service OS walks you through the offer, pricing and outreach. Our CLAUDE.md setup guide covers the repository side of the brief.
A project's threads write better when your brand is in front of them. In Brand Architect, Ophelia builds a knowledge base from your website or materials that any AI tool can read. Build your knowledge base →
