Setting Up a Claude Project You'll Actually Keep Using
5 min read
Most Claude Projects are created once, filled with enthusiasm, and never opened again. The cause is almost never the feature. It is a setup that asked too much on day one: forty uploaded files, a three-page instruction block, and no clear reason to open it on a Tuesday afternoon.
A project people keep using is small, specific, and tied to work that actually repeats. This article gives you a four-step setup, explains the limits and behaviors Anthropic documents, and shows how to tell when a project has gone stale. If you want the basic idea of why projects exist, the Complete Beginner's Guide to Claude introduces it first.
What a project is, by Anthropic's own description
Anthropic describes projects as self-contained workspaces with their own chat histories and knowledge bases. Each one has two things you control:
- Project knowledge: documents, text, code or other files you upload as background for Claude to use.
- Project instructions: guidance that tailors how Claude responds in that project, such as a formal tone or a specific role or industry.
Claude uses the project's knowledge to understand the context and background for every individual chat inside it. That is the payoff: you stop re-explaining, and a new chat starts already informed.
Facts that shape your setup
A handful of documented details change how you should build one.
- Projects are available to all users, including free accounts, which can create up to five.
- On paid plans, when project knowledge approaches the context limit, Claude switches on retrieval mode, which Anthropic says can expand capacity by up to ten times while keeping response quality.
- On Team and Enterprise plans, projects can be shared with members of your organization, with "Can view" and "Can edit" permission levels, and owners can turn project sharing off.
- Each project has its own memory, separate from your other projects.
The four steps
- 1
Name the recurring job, not the topic
"Weekly client report for the dental practice" is a job. "Marketing" is a topic. A job tells you what belongs inside and what does not, and it makes you open the project on the day the job comes round.
- 2
Upload only what you would hand a new colleague on day one
The brief, the style guide, two examples of work the client loved, the glossary of names. Not the archive. If you would not give it to a new hire, it does not belong in project knowledge.
- 3
Write instructions for how, not for what
Instructions should cover tone, format and what to do when something is missing: "Write in short paragraphs. If a number is not in the files, say it is missing instead of estimating." The what of each task goes in the chat.
- 4
Run one test chat before you trust it
Ask a question whose answer you already know and that depends on a file. "What did the client say about the footer colors?" If the answer is wrong or vague, fix the files or the instructions now, not three weeks in.
What good project instructions look like
Compare two versions of the same instructions.
A version that will not last: "You are an expert marketing assistant. Be professional, helpful and creative. Help with all client work."
A version that will: "This project is for one client, a dental practice. Audience: adults deciding on a first appointment. Tone: calm, plain, never salesy. Use the glossary file for treatment names. Keep emails under 150 words. If a request needs a price or an offer, ask me first; do not invent one."
The second works because it names the audience, the tone, the file to consult, a length limit and a rule for the exact moment Claude is most likely to make something up.
Tip
Write at least one instruction that begins with "If." Rules about what to do when something is missing, unclear or out of scope are the ones that prevent the worst mistakes.
One project per body of work
The idea is to keep different bodies of work apart, so their context never blends. A project for one client and a project for another beat a single project called "Clients." The same goes inside a company: a launch, a course, a research topic.
This is also where memory comes in. Because each project has its own memory, notes about one client do not appear in another. If a project holds something sensitive, you can also turn off "Use account memory" on the project's page so nothing from your general chats flows in.
If you are choosing between a project and a plain chat for a specific task, the decision test in ChatGPT Projects vs. a Regular Chat transfers well: will you come back to this, with context that has to carry over?
Signs a project has gone stale
Review a project the way you would review any shared folder. These are the signals.
A file in project knowledge contradicts a newer one (an old price list, a superseded brief).
The instructions mention a person, a product or a rule that no longer exists.
You keep correcting the same thing in chats, which means it belongs in the instructions.
You have not opened the project in a month, which usually means the job it was built for has moved on.
When a project goes stale, edit it before you stop using it. Fifteen minutes of cleanup beats starting over, and the repeated corrections you were making in chats are exactly the lines that should go into the instructions.
A note on long knowledge bases
More files is not automatically better. Retrieval mode helps Claude cope with a large knowledge base, but a pile of overlapping documents still gives it more chances to pick the wrong one. If two files answer the same question differently, Claude has no way to know which you prefer. Delete the loser.
Common mistake
Treating project knowledge as a place to keep everything just in case. It is context Claude will draw on, so everything in it should be something you want reflected in the work.
Official sources
- What are projects? | Claude Help Center
- Understanding Claude's personalization features | Claude Help Center
Checked on October 24, 2026. Features, plans and names change often, so the vendor's own pages are the final word.