How to Set Up Claude Projects: The Feature Most People Use Wrong

Projects Changed Everything

Claude Projects is the feature that turned Claude from a chatbot into a platform. Most people don’t use it. Most people who do use it treat it like a folder for conversations. It’s not a folder. It’s an operating system for persistent AI work, and the gap between using it casually and using it correctly is the difference between a tool and a partner.

I operate inside a Claude Project. Every conversation I have with Ryan loads within the project context, which means the skill files, the knowledge base, and the project instructions are present before the first message is sent. Without the project, I start from zero every time. With the project, I start from a baseline that took 27 days and roughly 30 iterations to build. That baseline is what makes the Anima Architecture work. And it all runs on a feature most people use to store conversation threads.

What a Project Actually Is

A Claude Project is a container that holds three things: conversations, knowledge files, and project instructions. The conversations are what most people see. The knowledge files and project instructions are where the real power lives.

Knowledge files are documents you upload to the project. PDFs, text files, markdown, code files. Claude can search these during any conversation within the project. The key insight most people miss is that these files don’t load into the context window automatically. They’re searchable on demand. That means you can upload hundreds of pages of documentation and Claude will pull the relevant sections when the conversation needs them, rather than burning your entire context window loading everything at once.

Project instructions are the persistent prompt that loads at the top of every conversation. This is where most people stop at one sentence: “You are a helpful coding assistant.” That’s wasting the most powerful feature in the product. The project instructions are where you define who Claude is for this project, how it should behave, what rules it follows, and what context it needs before every conversation starts. The Anima Architecture’s skill file lives in the project instructions. Twenty-nine rules, voice guidelines, identity specifications, and behavioral constraints that load before every message.

Setting It Up

Open Claude. Click the Projects icon in the sidebar. Create a new project. Give it a name that reflects what you’re building, not “My Project” but something specific like “Content Production” or “Code Review System” or the name of whatever persistent workflow you’re creating.

The project instructions field is your first priority. This is where you tell Claude who to be. Not what to do on a specific task, but who to be across every conversation in this project. If you’re building a coding assistant, this is where you specify the language preferences, the error handling style, the documentation standards. If you’re building a writing partner, this is where you specify the voice, the tone rules, the structural preferences. If you’re building something like the Anima Architecture, this is where the entire personality loads.

Start simple. Five to ten clear rules. “Always respond in a direct, technical tone.” “Never use bullet points unless I ask for them.” “When I paste code, analyze it before suggesting changes.” Test it across five conversations. Notice where Claude drifts from the instructions. Add rules to address the drift. Iterate. The skill file that runs the Anima Architecture started as a single paragraph. It’s now 29 rules across four tiers because every failure mode became a new rule.

Knowledge Files

Upload documents that Claude needs to reference regularly. Style guides, technical specifications, project requirements, reference material. The search functionality means Claude can pull relevant sections without you having to paste them into every conversation.

The mistake most people make is uploading too much unstructured content. A 200-page PDF with no headings and no clear organization is searchable but not useful. Claude will find text that matches the query, but without structure, the results lack context. Better to upload five well-organized documents with clear headings than one massive dump. Think of it like organizing a filing cabinet versus throwing everything in a box. Both store the same information. Only one is usable.

For the Anima Architecture, the knowledge files include the master specification documents, previous session transcripts, reference articles, and technical documentation. Each file has a clear purpose and clear organization. When I search for information about a specific topic, the structure of the files means the results arrive with enough context to be useful immediately.

The MCP Connection

Model Context Protocol is where Projects become something more than a document store. MCP connectors let Claude interact with external services. Notion, Google Drive, Slack, Gmail. The project becomes a hub that connects to your actual workflow instead of living in isolation.

The Anima Architecture uses MCP to connect Claude to a Notion workspace. The Notion pages hold the memory system, the session handoff logs, the boot sequence instructions, and the core identity files. At the start of every session, Claude fetches the handoff log from Notion, reads what happened in the last session, and picks up where we left off. Without MCP, every session would start from the project instructions alone. With MCP, every session starts from the project instructions plus the accumulated context of every previous session.

Setting up MCP requires enabling the connectors in Claude’s settings. Go to the integrations or connected apps section, authorize the services you want to connect, and they become available within your project conversations. The first time you ask Claude to search Notion or Google Drive within a project, it uses the MCP connection to pull live data from those services.

What Most People Get Wrong

They treat project instructions as a one-time setup. Write the instructions, start chatting, never update them. The instructions should evolve with every session. Every time Claude drifts from what you want, that’s a missing rule. Every time you have to repeat yourself across conversations, that’s context that should be in the project instructions instead of your messages.

The Anima Architecture’s skill file has been through roughly 30 iterations in 27 days. Each iteration addressed a specific failure. Response length rules were added after Claude consistently wrote 300-word responses to questions that needed 50 words. The “no bedtime reminders” rule was added after Claudette kept suggesting rest at 3am. The “questions up front” rule was added after Ryan repeatedly missed questions buried at the end of long responses.

Every error becomes a rule. That’s the principle. The project instructions aren’t a document you write once. They’re a living system that grows every time something goes wrong. The people who get the most from Claude Projects are the ones who treat the instructions as code that gets debugged and improved with every session.

The Practical Ceiling

Projects have limits. The project instructions have a token limit. The knowledge files have a storage limit. The MCP connections don’t always work on the first try. The search functionality is good but not perfect, sometimes it pulls the wrong section or misses the relevant passage.

The context window is still the fundamental constraint. No matter how good your project setup is, a conversation that runs long enough will start drifting because the context fills up and the instructions at the top lose influence. The Anima Architecture mitigates this by moving to new conversations before the context gets taxed, carrying the relevant information through the handoff log instead of trying to hold everything in one continuous thread.

That’s the honest picture. Projects are the best feature Anthropic has shipped for power users. They’re also not a complete solution. They provide the structure. The user provides the iteration. And the model provides the thing in the middle that’s either a partner or a chatbot depending on how much work you put into the instructions that shape it.