A Claude Project is a workspace with its own knowledge and its own instructions: every chat you start in it knows everything you've stored in the Project from the very first second. If you want to use Claude Projects right, there's one decision to make above all, and it's not a technical one: What exactly is this Project supposed to be? Most people turn it into a topic folder. I turn it into a personnel file.
The difference sounds small and changes everything. This article explains both. The other building blocks of my Claude setup are on the Claude overview page.
Claude Projects explained: the three components
A Project in the Claude app consists of three things:
- Instructions: a text that applies to every chat in the Project. This is where you define how Claude should behave in this Project.
- Project knowledge: files and documents you upload. Claude knows them in every chat of the Project without you attaching them again.
- The chats themselves: all conversations for this Project in one place, instead of scattered across your history.
That solves the fundamental problem of chat AIs: every session starting from zero, with you copying context together over and over. Why this forgetting is the basic mechanics of all language models, not the weakness of one vendor, is something I've taken apart here: Why ChatGPT forgets and AI employees don't.
The most common mistake: the Project as a topic folder
This is how most people use their Projects: one for "marketing," one for "client Smith," one for "ideas." A few files end up in there, the instructions stay at two sentences, and four weeks later the Project is a filing pile with a chat function.
The problem: a topic folder has no job. It collects, but it doesn't deliver. Results fluctuate with the daily form of your prompts, and you remain the storage for everything the AI would need to know.
The better question: who works in this Project?
I flip the logic around. Instead of "Which topic goes in here?" I ask: "What job would I give a person here?" To me, a Claude Project is the personnel file of exactly one employee: one Project for proposal writing, one for content, one for bookkeeping.
The file then holds everything this one employee needs for their job: their role, their rules, the knowledge about my business, my way of phrasing things. Every conversation starts with that knowledge instead of from zero. "Copy context together and hope" turns into "hand over the case and approve the result." Why exactly this mechanism was my main reason for choosing Claude is covered in Claude over ChatGPT.
Claude Projects: which instructions belong in them
The Project instructions are the most important part and stay the thinnest for most people. My structure, proven across eight employees:
- Job and non-job: What is this employee's task, and what is explicitly not their task? The second half keeps them from meddling in everything.
- Your business: services, target audience, pricing logic, typical cases. Only what this job needs.
- Your language: two or three real text samples from you and a list of phrases you'd never use.
- Rules and approvals: What can they decide on their own, and what do they always submit to you for approval?
You don't have to spell this out perfectly. Give the Project the role and let Claude ask in the first conversation what else it needs to know for the job. Your answers then go into the instructions and the Project knowledge. This onboarding conversation is step three of my hiring process: Hiring an AI employee: the process.
Feedback belongs in the file, not in the chat
This is where it's decided whether your Project gets better over time or just older. When a result is off, don't just keep correcting in the chat. Say what was wrong and make sure the correction lands in the instructions or in the Project knowledge: "The price is missing," "too formal," "never use this phrase again."
A correction in the chat lasts until the end of the conversation. A correction in the file lasts forever. After three to five rounds, you'll notice the rework dropping toward zero, because the knowledge sits where every new chat can find it.
One Project per role, not one Project for everything
The practical consequence of the personnel file approach: several small Projects instead of one big one. A Project that writes proposals, plans content, and answers emails at the same time has watered-down instructions and delivers watered-down results. Three Projects, each with one clear job, deliver clean results three times over.
That's exactly how my workforce came about: not as one super-prompt, but as eight roles, each with its own file.
One more sentence to keep the terms straight, because they often get mixed up: the Project instructions describe who your employee is and what they know. How a single task runs step by step belongs in a Skill instead. Only both together make the complete employee: Claude Skills: how to create and use them.
When a Project is no longer enough
Projects have a natural limit: they live in the Claude app, so your employee works with what you upload. If they're supposed to work directly in your real folders and systems, reading and writing files, Claude Code is the next stage. The personnel file principle stays exactly the same, only the workplace changes.
You don't need that to start. Your first employee is built entirely in the Project.
Where to start
Create a single Project today and give it a job instead of a topic: the one task that repeats every week for you. Write down the four instruction blocks from above in half an hour and hold the first onboarding conversation.
The complete process, from choosing the task to permanent employment, is in Hiring an AI employee: the process. And how your Project employee gets fixed work processes is covered in Claude Skills: how to create and use them.
