Claude Projects: what they are, how project knowledge works, and how to write good instructions

What are Claude Projects?

Claude Projects are workspaces in the Claude app that group related chats under one set of instructions and one set of files. Anthropic's help center describes them as "self-contained workspaces with their own chat histories and knowledge bases." Everything you upload to a project, and every instruction you set on it, applies to every chat you start inside it.

That solves the most common annoyance with chat tools: pasting the same background into every new conversation. Put your style guide, product docs or codebase notes into a project once, write a few lines of instructions, and each new chat starts with that context already in place.

Anthropic announced Projects in June 2024 for Pro and Team users. Since then the feature has opened to every plan, gained retrieval for large knowledge bases, and picked up per-project memory.

Which Claude plans include Projects?

All of them, with different limits. According to How can I create and manage projects?, "Projects are available to all users, including those with free Claude accounts," and "Free users can create a maximum of five projects."

The help center does not state a project cap for paid plans. For current plan names and prices, check claude.com/pricing.

How do you create a Claude project?

1. Open claude.ai/projects and click + New Project.
2. Give it a name and a description. Anthropic notes that Claude "will not have access to these details," so the description is only for you and your teammates.
3. On Team or Enterprise, choose whether the project is private or shared with your organization.
4. Add files with the + button on the project page.
5. Click Set project instructions, write them, and choose Save instructions.

You can also move an existing chat into a project from that chat's menu, star a project to pin it in the side panel, archive it, or delete it from the three-dot menu. An archived project keeps its knowledge and sharing settings, but you have to unarchive it before you can delete it.

What is project knowledge?

Project knowledge is the set of files you add to a project. Claude uses them as background in every chat inside that project. You can upload documents, text and code, and paste text snippets directly.

Anthropic's file upload article gives these limits for projects:

• File size: 30MB per file (chats allow 500MB per file).
• Number of files: "Unlimited, but total content must fit within Claude's context window."
• Processing: "Text extraction only (except for multimodal PDFs)," so images inside most documents are not read in project knowledge.
• File types: the article's general upload list is PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSX, plus JPEG, PNG, GIF and WebP images. XLSX needs code execution and file creation turned on.

Because extracted text is what counts, a 25MB PDF full of scanned images may contribute little, while a 2MB text export of your docs may fill a lot of the space. Clean, text-heavy files work best.

How does project knowledge search (RAG) work?

On paid plans, once your knowledge gets large, Claude stops loading all of it into every chat and searches it instead. Anthropic's article on RAG for projects explains the switch:

• "When your project knowledge approaches the context window limit, Claude will automatically enable RAG mode." You don't turn it on.
• RAG can "expand your project's capacity by up to 10x while maintaining quality responses."
• Claude uses a "project knowledge search tool" to pull only the relevant parts of your files for each question. You can see it run in the chat.
• The project shows a visual indicator when RAG is on, and if your knowledge shrinks below the threshold, "Claude can automatically convert back to context-based processing."

RAG changes how you should prepare files. In full-context mode, Claude sees everything at once. In RAG mode, it sees only the passages its search finds, so:

• Use descriptive headings and file names, since they help the right passage match a question.
• Split one giant document into topic files rather than one 400-page export.
• Ask specific questions that use the words in your documents. "What does the refund policy say about annual plans?" retrieves better than "tell me about refunds."
• If an answer seems to miss something, ask Claude to search the project knowledge for it by name.

How do you write good project instructions?

Project instructions tell Claude how to behave in every chat in the project. Anthropic's help center gives the classic use: "instructing Claude to use a more formal tone or answer questions from the perspective of a specific role or industry." In practice, the best instructions are short, specific and about the work, not about Claude.

A good set covers four things:

1. Who it's for. The audience and their level.
2. What the knowledge is. Which files matter most and how to treat them.
3. How to answer. Format, length, tone, and what to do when unsure.
4. What not to do. The one or two mistakes you keep correcting.

Example: a product docs project

You are helping the support team answer customer questions about Acme Billing.

- The files in this project are the current help center and pricing docs.
  Treat them as the source of truth. If they don't answer a question, say so.
- Answer in plain language a non-technical customer understands.
- Keep replies under 150 words unless asked for more.
- Quote the help center article title you used at the end of each answer.
- Never promise refunds or discounts; point to the refund policy instead.

Example: a codebase notes project

This project holds architecture notes and API docs for our Django backend.

- Assume Python 3.12 and Django 5 unless a file says otherwise.
- When suggesting code, match the patterns in style-guide.md.
- Prefer small, reviewable changes and explain the trade-off in one sentence.
- If a question needs code you can't see, ask for the file instead of guessing.

Example: a writing project

You are my editor for a weekly newsletter about personal finance.

- voice.md describes my tone. Follow it closely.
- past-issues.pdf shows the format: one idea, three short sections, one takeaway.
- Edit for clarity first, then length. Flag claims that need a source.
- Don't add emojis or exclamation marks.

Keep instructions focused on what changes Claude's answers. Long lists of general advice ("be helpful, be accurate") use space without changing much. If you find yourself writing a step-by-step procedure, that may belong in a skill instead, which the next sections cover. For ready-made openings, see these Claude prompts and prompt templates.

How does sharing work on Team and Enterprise?

Sharing is for Team and Enterprise plans only. Per Manage project visibility and sharing:

• Public means "Everyone in your organization can view and use the project." Private means only invited members can.
• Invited members get Can view (see the knowledge and instructions and chat in the project) or Can edit (change instructions and knowledge and manage members).
• Chats inside a project stay private unless the chat owner shares them.
• Owners can turn off project sharing for the whole organization (which also makes public projects private), and sharing with a group is an Enterprise feature in beta.

A shared project is a good way to give a team one agreed set of instructions and reference files, so everyone gets answers from the same docs.

Do projects have memory?

Yes, separately. Each project has its own memory, kept apart from your other projects and from chats outside projects. The help center says memory is on by default on Free, Pro and Max, and an owner must enable it on Team and Enterprise. You can stop a project from using your account memory, or turn memory off for a single chat.

Memory is not the same as knowledge. Knowledge is the files you choose to add. Memory is what Claude picks up from your conversations over time. If something must always be true, put it in the instructions or a knowledge file rather than relying on memory.

Projects vs Skills vs custom instructions vs Claude Code

The Claude app has several places to put instructions, and they overlap. Anthropic's personalization article separates account-wide instructions, project instructions and skills.

Some rules of thumb:

• "I always want this" goes in Instructions for Claude.
• "For this client, product or book" is a project.
• "Whenever I do this kind of task" is a skill. A skill loads only when needed, so it works across projects without copying it into each one.
• "In this repo" belongs in CLAUDE.md, which Claude Code reads at the start of each session.

Anthropic is also rolling out a new version of Projects, starting at claude.ai/code, in public beta on Pro and Max. It works differently: one conversation that runs related tasks as parallel cloud sessions sharing repositories, instructions and memory. The Claude Code docs say chat and Cowork projects "keep working as they do today until the redesigned experience reaches them." This page covers those current chat projects.

Keep your instructions somewhere you own

Good project instructions get reused. The support-tone block from one project ends up in the next, and the codebase rules get copied between workspaces. Kipt keeps prompts like these as Markdown in a private GitHub repo you own, with Cmd+K search and blanks such as {{product:Acme Billing}} that you fill in when you copy. When an instruction turns into a procedure, the free SKILL.md generator turns it into a SKILL.md that works in any chat, not just one project. It runs in your browser and needs no sign-up.