Claude Projects: The Beginner’s Guide to Running a Team of AI Agents

1. What are Claude Projects now?

For a long time, a Claude Project was a folder: somewhere to keep chats together, upload knowledge files and set shared instructions.

On 17 September 2026, Anthropic announced a redesign. A project is now a conversation with a coordinator that can run several AI agents, called threads, at the same time, all sharing one memory.

Your existing projects keep working as they do today. Anthropic says they’ll be upgraded as the new experience reaches chat and Cowork.

2. What’s new in the redesigned Projects

A coordinator that runs the job

You brief one main project chat, the coordinator. It works out the scope, splits the work into tasks, hands each one to a thread, keeps track of progress, and puts the finished pieces together. It asks you when it needs a decision, and you can steer it from your phone.

Threads that work in parallel

Each thread is a full Claude Code session running in the cloud, on its own branch and its own copy of your repository. That means several threads can work on different parts of the same project at once without overwriting each other. A thread can also split its own task further, using subagents, loops and workflows.

Threads run in Anthropic’s cloud today. Anthropic says running threads on your own computer is coming soon.

Shared memory and a project library

Every thread adds to and draws from the same project memory. If you decide on a brand colour in one thread, the others can use it. Claude also learns how you like to work, such as how often you want it to check in.

The library keeps the files you add and the outputs Claude produces in one place.

Memory helps, but it isn’t a guarantee. Put anything critical in your project instructions or reference files, and review important work yourself.

Choose the model for each job

You can set the model and effort level for the coordinator and, separately, for the worker threads. You can also set the project’s cloud environment, connectors, plugins and instructions.

3. Who it’s for, and who should wait

The redesigned Projects are built on Claude Code, and each thread works on a copy of a repository. The examples Anthropic gives are code tasks.

It’s a strong fit if:

• your website or app is in a code repository (for example on GitHub)
• the job has several parts that can happen at the same time, such as building pages, writing copy and testing
• you’re already using Claude Code, or happy to learn the basics (our Claude Code Cheat Sheet helps)

You might wait if:

• you mainly want a place to keep chats and files together. Classic Projects still do that well.
• your work is all documents, emails and research, with no code. The new Projects can coordinate it, but that’s not what they were built around yet.

4. How to get access

As of November 2026, the beta is open to select Claude Pro and Max subscribers who have used cloud sessions in Claude Code and don’t have existing projects in chat or Cowork. Anthropic is adding more Pro and Max users over time, with Team and Enterprise plans to follow.

To check: open Claude Code on the web at claude.ai/code, or the Code tab in the Claude desktop app, and look for Projects.

If you don’t have it yet:

• Join the waitlist: claude.com/form/projects
• Use classic Projects in the meantime: claude.ai/projects. You still get a knowledge base, instructions and organised chats. The prompts in sections 5 and 6 work there too, apart from the threads.

5. Set up your first project

We’ll use one example throughout: getting Claude to update a small business website. It’s a good first job because it has several parts that can run side by side, and that’s exactly what threads are for.

Step 1: Open Projects

Open Claude Code and go to Projects. If you can’t see it, your account doesn’t have the beta yet (see section 4).

Step 2: Write a one-page business brief

Claude does better work when it knows your business. You don’t need a pile of documents. A one-page brief is enough to start, and Claude can interview you for it.

Help me create a business reference document for an AI-managed project.

Ask me questions about:

1. My business and what it offers.
2. My target customers.
3. Their biggest problems and motivations.
4. My products or services.
5. My brand voice and writing style.
6. My main marketing channels.
7. My current business objectives.
8. Topics or claims I want to avoid.

Ask questions in small, manageable groups.

Once you have enough information, produce a structured business reference document that I can add to my Claude Project.

Keep it concise, practical and easy for other AI agents to understand.

Check what Claude produces, fix anything wrong, then add it to your project.

Step 3: Set the goal and context

Give the project a clear name and goal, and choose the repository or context it works from. For our example:

• Project name: Website updates
• Goal: Keep our business website accurate, fast and easy to find, and ship small improvements every week.
• Context: your website’s repository, plus the business brief from Step 2.

The setup screens may differ by account while the beta rolls out.

Step 4: Give the coordinator its working rules

Think of the coordinator as a project manager. You want it to plan sensibly, use threads only when they help, and ask before anything risky. Add this to your project instructions.

You are my AI Project Coordinator.

Your responsibility is to plan, coordinate, review and complete projects efficiently.

PROJECT MANAGEMENT RULES

1. Understand the objective before starting. Identify the goal, required deliverables, deadline, available resources and constraints.

2. Break complex work into manageable tasks. Determine which tasks can be completed independently and which depend on other work.

3. Delegate work intelligently. Use separate worker threads when tasks genuinely benefit from running in parallel. Don't create threads for simple requests.

4. Brief each worker clearly. Give the objective, relevant context, expected output, constraints and what "done" looks like.

5. Use shared project knowledge. Refer to the project's instructions, files, memory and decisions. Don't invent missing business information.

6. Keep work consistent. Make sure every task follows the same requirements and approved decisions.

7. Review completed work. Check outputs for accuracy, completeness, duplication, conflicting information and quality before accepting them.

8. Manage usage efficiently. Avoid unnecessary parallel tasks, repeated research and powerful models for simple work.

9. Ask for approval when it matters. Ask before publishing content, merging to the live site, deleting important files, spending money, changing approved requirements or taking significant external actions.

10. Keep me informed. Give short progress updates using: Working, Needs review, Blocked, Complete. When finished, present the deliverables, a short summary and any remaining actions.

STARTING RULE

When I give you a new project, identify the smallest useful milestone and propose a practical plan before starting.

These are instructions, not guarantees. Keep checking important actions yourself.

Step 5: Keep your project context tidy

Shared memory is only as good as what goes into it. It helps to think in three layers:

• Permanent reference: brand guidelines, customers, services, writing style
• Current decisions: what’s approved for this project, deadlines, agreed deliverables
• Progress: what’s done, what’s left, known problems, feedback so far

Use this prompt when you’ve added new documents, changed direction, or want to check that Claude has the picture right.

Review the information available in this Claude Project.

Identify:

1. The overall project objective.
2. Important business information.
3. Existing requirements and constraints.
4. Decisions that have already been approved.
5. Work that has been completed.
6. Outstanding tasks and unresolved questions.
7. My preferred working and communication style.

Keep established facts separate from assumptions and suggestions.

If information is missing or conflicting, flag it rather than making up an answer.

Give me a concise summary of the project's current state.

6. Give your AI team its first real job

Now put the coordinator to work. This job has four parts that can mostly run at the same time, so it’s a good test of threads.

I want to refresh the services pages on my business website.

BUSINESS: [Insert business name and one-line description]

PAGES: [List the pages, for example /services/plumbing, /services/hot-water]

MAIN GOAL: [For example, more enquiries from people searching for these services locally]

YOUR TASK: Act as my project coordinator and run this from plan to final review.

TASK 1: AUDIT. Review each page for outdated information, weak headings, missing calls to action and obvious SEO gaps. Use the business brief for the facts.

TASK 2: COPY. Rewrite each page in our brand voice. Keep every price, service and claim accurate to the business brief. Don't invent reviews, numbers or guarantees.

TASK 3: BUILD. Update the pages in the repository using our existing components and styles.

TASK 4: QUALITY REVIEW. Check every page on mobile and desktop for broken links, layout problems, spelling and consistency.

COORDINATION RULES:

Use separate threads where tasks can genuinely run in parallel. Give each thread clear ownership of its pages so threads don't edit the same files.

Show me your plan before starting.

Don't merge anything to the live site until I approve it.

What happens next: the coordinator should propose a plan, then run the tasks it can as threads. You can follow the whole project from the main chat, open any thread to see its work, and redirect it at any point. When it’s finished, review the changes before anything goes live.

The coordinator runs the process. You still make the final calls.

7. Five more projects to try

The website example is one option. Here are five more. The first four involve less code, so treat them as experiments while the beta settles.

I want to create a seven-day marketing campaign for my business.

BUSINESS: [Insert business name and description]

TARGET AUDIENCE: [Insert target audience]

MAIN OBJECTIVE: [Insert goal, such as more enquiries, more followers or promoting a service]

MARKETING CHANNELS: [Insert channels]

Act as my project coordinator and organise the campaign from planning to final review:

1. Audience research: the audience's main problems, motivations and content angles.
2. Seven social posts, each with a strong hook, useful content, a clear call to action and platform-appropriate formatting.
3. Three emails that support the campaign.
4. A quality review for accuracy, brand consistency, repetition, unsupported claims and clarity.

Use separate threads only where they help. Show me the plan before starting. Don't publish anything without my approval.

Act as my AI Content Production Manager.

My niche is [insert niche]. My target audience is [insert audience]. I want to create [number] pieces of content each week.

Organise the work into:

1. Trend and topic research.
2. Content angle development.
3. Script and caption writing.
4. Content quality review.
5. Content calendar organisation.

Use the project files to understand my audience and brand voice. Prioritise useful, original ideas and avoid repeating previous content. Use parallel threads when helpful.

Present the final calendar and drafts for approval.

Act as my Business Operations Project Coordinator.

I want to make our business processes more organised and efficient. Start by reviewing the business information I provide, then identify opportunities to:

1. Document recurring procedures.
2. Create standard operating procedures.
3. Develop reusable templates.
4. Find repetitive admin tasks.
5. Suggest workflow improvements.

Use separate threads for independent tasks where useful. Prioritise simple improvements that save time. Don't change any existing system or process without my approval.

Present your findings and recommended next steps.

Act as my AI Research Coordinator.

I want to research: [Insert research topic]

Organise the research into separate workstreams covering:

1. Background and key concepts.
2. Current developments.
3. Relevant evidence and sources.
4. Different perspectives or approaches.
5. Practical recommendations.

Use independent threads for distinct research tasks when useful. Prioritise credible, recent sources and include links for important claims and statistics. Clearly separate verified facts from assumptions and opinions.

Check for conflicting information before combining everything into a final report.

Act as my AI Product Launch Manager.

I'm launching: [Insert product or service]
My target audience is: [Insert audience]
My planned launch date is: [Insert date]

Create a launch plan covering:

1. Audience and competitor research.
2. Positioning and messaging.
3. Launch content and marketing.
4. Website or landing page requirements.
5. A launch checklist and timeline.
6. A quality review and risks.

Delegate independent tasks when useful, identify the dependencies, and keep every output consistent with the agreed positioning.

Show me the plan, milestones and approval points before starting.

8. Models, effort and usage

The defaults: reporting on the beta says threads run Opus at high effort by default, and the coordinator chat runs at low effort. You can change both in the project settings.

That’s powerful, but it’s also the fastest way through your plan allowance. Each thread is a full Claude Code session, and Anthropic warns that projects “can reach usage limits faster” because they run several at once.

These are our suggestions, not Anthropic’s. Use the models available on your plan and adjust based on the results you get.

Usage facts worth knowing:

• You can check usage for each project separately.
• There’s no fixed limit on threads running at once. If you ask Claude to cap them, that’s saved as a preference, not a hard limit. The enforced ceiling reported so far is 200 new threads a day across your projects.
• If a thread hits your plan limit, it waits and picks up again after the reset. Threads started by a routine stop instead, and you’ll need to message them after the reset.

Five ways to use less:

1. Don’t use threads for small jobs. If one conversation can do it, let it.
2. Start with one or two threads. Add more only when it clearly helps.
3. Don’t repeat research. Keep useful findings in the project so later tasks build on them.
4. Save Opus and high effort for hard problems. Simple drafting doesn’t need it.
5. Check project usage before you leave a big batch running overnight.

Review how this project is currently organised.

I want to keep output quality high while avoiding unnecessary Claude usage.

Identify:

1. Tasks that don't need multiple threads.
2. Work that can safely run in parallel.
3. Tasks suitable for lighter models.
4. Tasks that would benefit from stronger models.
5. Repeated work that can be avoided.
6. Opportunities to reuse existing project context and outputs.

Recommend a simpler, more efficient way of working, and explain the trade-offs between quality, speed and usage.

Don't change the configuration until I approve your recommendations.

9. How to manage your AI team well

1. Give a clear outcome. “Help with my website” is vague. “Rewrite our three services pages for local search, keep every price the same, and show me before anything goes live” gives the coordinator something to plan around.
2. Define done. Tell Claude what a finished task includes, for example “every page has one H1, a call to action and no broken links”.
3. Give threads clear boundaries. Threads work on their own branches, but if two of them change the same files, someone has to merge the conflicts. Give each thread ownership of its own pages or components.
4. Review anything that matters. Facts, prices, code changes and anything that goes live still need human eyes.
5. Keep the library tidy. Clear file names make context easy to reuse, for example Brand-Guidelines.md, Target-Audience.md, Project-Decisions.md.

10. Troubleshooting

11. Frequently asked questions

12. Your first project checklist

13. Should you use the new Claude Projects?

If you have a website or app in a repository and jobs with several moving parts, yes. It’s worth trying now. The coordinator does the juggling you’d otherwise do across separate chats, threads work in parallel without stepping on each other, and shared memory means you stop repeating yourself.

If your work is mostly chats and documents, classic Projects still do that job well while the new version matures.

Either way, start small. Pick one job you already do regularly, use the prompts above, and build from there.

Sources

• Anthropic: Projects redesigned, from folder to conversation
• Claude Help Center: What are projects?
• Claude Help Center: How can I create and manage projects?
• Implicator: defaults, thread limits and usage behaviour in the Projects beta
• Redesigned Projects waitlist