How to Use Claude AI Projects to Organize Conversations and Knowledge

Anthropic launched Projects with a 200,000-token knowledge base per workspace, roughly 500 pages of reference material that persists across every conversation you start in that project. North Highland, the management consultancy, reported completing tasks up to 5x faster after grounding Claude in their internal knowledge through the feature. That speed gain captures the core value: stop re-uploading your style guide, codebase documentation, or research notes every time you open a new chat.

Claude AI Projects are persistent workspaces that group related conversations with a shared knowledge base, custom instructions, and conversation history. Every chat within a Project inherits the same context automatically. Instead of pasting your company's voice guidelines into every conversation about marketing copy, you upload them once to the Project's knowledge base, set your tone preferences in the custom instructions, and every new conversation starts with that context already loaded.

The feature works across four tiers. Free users can create up to five Projects, which is enough to separate personal research, a side project, and a couple of work contexts. Pro, Max, Team, and Enterprise plans get unlimited Projects with additional capabilities: RAG mode that expands knowledge base capacity by up to 10x, separate memory summaries per project, and organizational sharing controls.

Projects also serve as a natural organizing layer for anyone who has accumulated hundreds of Claude conversations. Rather than scrolling through a chronological list of chats trying to find the one where you discussed your database schema, you create a "Backend Architecture" Project. Every conversation about schemas, queries, and migrations lives there, and Claude already knows your stack when you start a new thread.

The creation process takes about two minutes, but the decisions you make during setup determine how useful the Project becomes over the following months.

Step 1: Open the Projects panel. Navigate to claude.ai/projects or hover over the left sidebar in Claude's interface and click "Projects." You will see any existing projects listed here, along with a "+ New Project" button in the upper right corner.

Step 2: Name the project descriptively. Click "+ New Project" and enter a name that your future self will recognize. "Marketing Q3 Campaign Copy" is better than "Marketing." "Client Portal API Docs" is better than "Work Stuff." Claude does not read the project name or description for context, so the name is purely for your own organization.

Step 3: Set the visibility. On Team and Enterprise plans, you choose whether the project stays private or becomes visible to your organization. Free and Pro users skip this step since their projects are private by default.

Step 4: Add your knowledge base. Click the "+" button inside the project to upload documents. You can add PDFs, text files, code files, and other document types. Start with the documents you find yourself re-uploading most often: style guides, technical specs, product requirements, or research summaries. You do not need to upload everything at once. Add files as your project evolves.

Step 5: Write your custom instructions. Click "Set project instructions" to define how Claude should behave in this project. These instructions apply to every conversation within the project. Start with the most important constraints: "Always respond in British English," "Cite specific sections of the uploaded API docs when answering technical questions," or "Assume I'm a senior developer and skip introductory explanations."

Once these five steps are complete, every new conversation you start in the project will have your knowledge base and instructions loaded from the first message. You can also move existing conversations into a project by selecting them from your chat history.

The knowledge base is where Projects either save you significant time or become a cluttered document dump. The difference comes down to what you upload and how you structure it.

Upload living documents, not static snapshots. The most effective knowledge bases contain documents that change over time and need to be re-referenced frequently: coding standards, brand guidelines, product feature lists, research bibliographies, and meeting note summaries. If a document is a one-time reference you will never revisit, it belongs in a regular conversation upload, not a Project knowledge base.

Keep the context window in mind. Each Project has a 200,000-token knowledge base on the base tier. That sounds generous, but a 50-page technical specification can consume 15,000 to 20,000 tokens depending on formatting. Upload the sections Claude needs, not the entire document. If you are working with a codebase, upload the architecture overview and the modules you are actively working on rather than every file in the repository.

Paid plans get RAG mode. When your knowledge base approaches capacity on Pro, Max, Team, or Enterprise plans, Claude automatically enables Retrieval Augmented Generation mode. RAG expands your effective capacity by up to 10x while maintaining response quality. Instead of loading every document into the context window at once, Claude retrieves the most relevant sections based on your query. This means you can upload substantially more material without worrying about hitting the ceiling.

Structure your uploads by purpose. Create a naming convention that tells you what each file does at a glance. Prefix reference documents with "REF," templates with "TMPL," and examples with "EX." When your knowledge base grows beyond 10 files, this organization prevents you from uploading duplicates or losing track of what a file contains.

For teams that need a persistent, versioned knowledge base accessible to tools beyond Claude, platforms like Fast.io provide shared workspaces with built-in document indexing. You upload files once, and any tool with MCP access, including Claude through its MCP integration, can query them. This avoids maintaining separate copies of the same reference documents across different AI tools.

Stop re-uploading reference documents across AI tools

Fast.io gives you 50GB of free workspace storage with built-in document indexing and an MCP endpoint at mcp.fast.io. Upload your reference materials once, share them with your team, and access them from Claude or any other AI tool. No credit card, no trial, no expiration.

Sign Up Free

Custom instructions are the behavioral layer of a Project. They tell Claude how to respond, what to prioritize, and what to avoid. The difference between a useful Project and one that feels like plain Claude is almost always the quality of the instructions.

Start with constraints, not capabilities. Telling Claude what not to do is often more effective than describing what you want. "Never use bullet points in blog post drafts" is clearer and more enforceable than "write in flowing prose paragraphs." "Do not suggest solutions that require a paid third-party service" prevents entire categories of unhelpful answers.

Be specific about format. If you want code examples in TypeScript, say so. If you want responses under 500 words, say so. If you want Claude to ask clarifying questions before starting long tasks, say so. Vague instructions like "be helpful" or "be creative" accomplish nothing because Claude defaults to those behaviors anyway.

Layer your instructions with profile preferences. Claude stacks profile preferences and project instructions, with profile preferences loading first. This means you do not need to repeat universal rules in every project. Set your general preferences (language, tone, response length) in your profile, then use project instructions for context-specific guidance. A software engineering project might include: "This project uses Python 3.12, FastAPI, and PostgreSQL. Follow PEP 8. Assume the reader is familiar with async/await patterns."

Include examples of good output. One concrete example of the response format you want teaches Claude more than three paragraphs of abstract description. Paste a sample answer, label it "Example of the format I expect," and Claude will pattern-match against it consistently.

Update instructions as your project evolves. The worst project instructions are the ones you wrote in the first five minutes and never touched again. After a week of conversations, you will notice patterns: Claude keeps making the same mistake, or you keep requesting the same adjustment. Add those corrections to the instructions so they apply automatically going forward.

Claude Cowork brings Projects to your desktop with a critical addition: direct access to your local file system. While claude.ai Projects work through uploaded documents, Cowork projects can reference entire folders on your computer, watch for changes, and maintain memory across sessions.

Setting up a Cowork project gives you three options. You can start from scratch with a new folder, import an existing claude.ai project (which pulls over files and instructions), or point Cowork at an existing folder on your machine. The third option is the most practical for active codebases or document repositories. Point Cowork at your project directory, add instructions, and Claude can read, reference, and work with every file in that folder without manual uploads.

Memory is where Cowork projects diverge from their web counterparts most significantly. Cowork automatically enables per-project memory, meaning Claude remembers context from previous tasks and applies it to new ones. If you ran a task last Tuesday that restructured your CSS files, Claude remembers that structure when you ask about styling decisions today. This persistence eliminates the "starting from scratch" problem that plagues even well-configured web projects.

Cowork projects also support scheduled tasks. You can set up recurring activities within a project: generating a weekly summary of changes in a folder, running a daily check against a style guide, or producing a status update from your project notes. These automations run within the project's context, so they have access to the same instructions, memory, and files as your interactive sessions.

One practical workflow: set up a Cowork project that watches your documentation folder, with instructions to flag inconsistencies and outdated references. When your codebase changes, Claude catches documentation drift without you having to remember to check. Pair this with a shared workspace on a platform like Fast.io for the files that need to travel beyond your local machine, and you get both local convenience and team-wide access.

Cowork projects currently require a paid plan (Pro, Max, Team, or Enterprise) and run on macOS or Windows through the Claude Desktop application.

Team and Enterprise plans unlock project sharing with two permission levels that control what collaborators can do.

Can Use lets teammates view the project's knowledge base and instructions and start their own conversations within the project. They cannot modify the knowledge base or change the instructions. This is the right permission for most team members: they benefit from the shared context without accidentally overwriting the project's configuration.

Can Edit grants full access to modify instructions, add or remove knowledge base documents, and manage member settings. Reserve this for project owners and leads who need to maintain the project's configuration.

Sharing works through three channels. You can invite individual members by email, add several members in bulk, or set the project to organization-wide visibility (if your administrator has enabled this). Once shared, a project appears in each member's "Shared with me" tab, and members receive email notifications when they are added.

One important behavior to know: when you archive a project, all sharing permissions reset to private and the sharing context is wiped for security purposes. If you reactivate the project later, you will need to re-invite collaborators. This is intentional. It prevents stale shared projects from lingering with outdated access lists.

Starring helps manage the growing list of projects. Star your most-used projects to pin them at the top of your project list. With unlimited projects on paid plans, starring becomes essential once you pass a dozen active workspaces.

For teams that need to share reference documents beyond Claude, whether with stakeholders who are not on your Claude plan, with other AI tools, or with automated workflows, exporting the knowledge base files to a shared workspace keeps everything accessible. Fast.io workspaces handle this well: upload the same reference documents your Claude Project uses, set granular permissions per folder, and anyone on the team can access them through the web UI or API, regardless of which AI tools they use.