Claude Projects are workspaces within claude.ai where you group instructions, files and chats around one topic. A Claude project saves repetitive work: you set the context once, and every new chat in that project uses it automatically.
Creating a project is easy; the skill lies in how you build it. A carelessly built project gives inconsistent output, much like an ad hoc prompt. A well-built project produces consistent, predictable answers. Below I explain which choices make the difference, based on what consistently works in my own work and where things go wrong in practice.
What is a Claude project?
A Claude project is a self-contained workspace in Claude where all chats share the same instructions and knowledge base. You upload the relevant files once, you set out how Claude should respond, and from then on every chat within that project starts with that context already loaded.
A project only pays off when the instructions in it are right. How to build them is covered in the LearnLLM Claude Course, where project instructions and system prompts become a fixed part of your way of working.
The three core components are project instructions (comparable to a system prompt), project knowledge (the files Claude can consult) and the chats that take place within the project. What you say in one chat is not visible to other chats in the same project, except for information you explicitly add to the project knowledge.
- Role
- Context
- Way of working
- Hard rules
If you work with literature and datasets rather than client files, you can also use Claude Science for research as a separate working environment alongside projects.
Since February 2026, projects have been available on all plans, including the free plan. There are limits, though: free users can create up to five projects, and the RAG extension for large knowledge bases is only available on paid plans (Pro, Max, Ultra, Team and Enterprise). Sharing with colleagues only works on Team and Enterprise plans. The Claude models available in your chats also vary by plan.
When is a Claude project worth it?
A Claude project is worth it as soon as you have a recurring task that needs the same context every time. A one-off question does not deserve a project. A type of work you do three or four times a week, with the same background documents or instructions each time, does.
In my own work I use projects for recurring SEO and content tasks. The project for LearnLLM content, for example, contains the tone of voice, audience description, competitor analysis and structure rules that are relevant to every article. I do not have to supply them again with every new brief. The project for SERP analyses contains a fixed method and examples of earlier analyses, so the output stays comparable across different keywords.
When not to create a project: for one-off questions, for experiments where you actually want to test different angles, or for ad hoc analyses without a recurring structure. A project adds overhead in the form of setup and maintenance. For one-off work, that is wasted time.
The rule of thumb I use: will I do this type of work at least five times within three months? If so, a project is worth the effort. If not, a good prompt will do.
How do you create a Claude project?
Creating a Claude project takes less than two minutes. The setup itself is trivial; the real work is in the content.
Open claude.ai and click Projects in the left-hand sidebar, or go straight to claude.ai/projects. Click ‘+ New Project’ at the top right. Give the project a name and description. The description is meant for you, to tell projects apart; Claude does not use this text as an instruction. On Team and Enterprise plans, you choose in the same step whether the project stays private or is visible to your organisation.
Once it is created, you see three sections you can fill in: project instructions, project knowledge and the chat area. Start with the instructions and only then add files to the knowledge base. A project without instructions works much like a normal chat with attachments; it is the instructions that give the project direction.
What do you put in the project instructions?
Good project instructions tell Claude three things: which role it takes on, which context it works in, and which rules or restrictions apply to the output. Vague instructions such as ‘help me with content’ produce generic output. A specific version such as ‘take on the role of a senior content strategist with 15+ years of experience in B2B SaaS, assess texts for readability and fit with the audience, and never use exaggerated sales language’ steers the output down to the detail.
A workable structure I use in almost all my projects consists of four blocks. First a role description with a realistic level of experience. Then a short description of the context: who the work is for, in which sector, with what goal. Next the way of working: which steps Claude follows, in which order, and which interim checks it carries out. The last block contains the hard rules Claude must not deviate from, such as banned phrases, required structural elements or fixed formatting choices.
A common mistake is overloading the instructions. Instructions running to thousands of words do not automatically produce better output; often the opposite. What recurs in my projects is fixed roles, fixed ways of working and hard rules. I deliberately keep two things out of the instructions: examples of earlier output (they belong in the project knowledge) and one-off briefs (they belong in the chat itself).
How do you use project knowledge effectively?
Project knowledge is where you upload documents Claude needs to consult during every chat in the project. Each file has a limit of 30 MB and the number of files is technically unlimited, but there is a token limit that determines the practical capacity.
On paid plans, the context window is 1 million tokens for current models such as Opus 5.5, Sonnet 5.5 and Haiku 5.5; older models have 500,000 or 200,000 tokens, where 200,000 tokens is roughly 500 pages of text. As soon as the combined content of your project knowledge approaches that limit, Claude automatically switches to RAG mode (retrieval-augmented generation). In RAG mode, not all content is loaded into memory at once; for each question, Claude looks up the relevant chunks in the uploaded files. This raises capacity roughly tenfold, without affecting the quality of the answers. On paid plans, this switch happens automatically.
Which files you upload matters more than how many. A good starting point is to clean up all files before uploading. A clean document without footer noise, duplicate headings or outdated versions gives better retrieval than the same content in a messy PDF.
It also helps to use recognisable file names. Claude partly relies on the file name to decide which file is relevant to which question. ‘Tone-of-voice.md’ works better than ‘document_v3_final_FINAL.docx’.
Grouping files by topic pays off too. Five well-structured files on one topic give sharper results than fifty loose files that all touch on the topic a little. And actively remove old files: outdated versions in the project knowledge lead to inconsistent output, because Claude sometimes draws on the old version and sometimes on the new one.
A useful feature of project knowledge is that the content is cached. When you consult the same files several times in different chats within the project, only the new part counts towards your usage limits. That makes projects more efficient than uploading files again and again in separate chats.
How do you share a Claude project with your team?
Sharing a Claude project is only possible on Team and Enterprise plans. On Pro or the free plan, the project stays private to whoever created it. Sharing on business plans works at two levels: organisation-wide sharing and individual sharing.
To share a project, open it and click ‘Share project’ to the right of the project name. There you add colleagues by name or email address, or paste a list of addresses to share in bulk. You then choose a role for each person. ‘Can view’ gives access to the content, instructions and knowledge and the ability to chat in the project, but without editing rights. ‘Can edit’ gives full control, including changing the instructions, adding or removing files and managing members.
Chats within a shared project are not visible to other members by default. What is shared is the workspace (instructions and knowledge), not automatically the conversations you have in it. If you want to share a specific chat with a colleague, use the ‘Share’ button within that chat to create a snapshot. That distinction prevents draft conversations or experiments from becoming visible to the whole team by accident.
Besides sharing projects, you can also collaborate through Claude Tag, where colleagues bring Claude straight into a Slack channel.
For organisations that want to work with AI in a uniform way, this is the main advantage of Team and Enterprise plans. One person sets up the project according to the company standard, shares it with the department, and from then on everyone works with the same instructions and the same knowledge base.
How do you move a chat into a project?
You can move existing chats into a project afterwards. That is useful when you realise an ad hoc chat actually belongs in an ongoing workstream, or when you want to set up a new project based on earlier conversations you have already had.
For a single chat, click the arrow next to the chat name in the sidebar and choose ‘Add to project’. In the pop-up window, find the right project and click it to move the chat. For several chats at once, use bulk selection in the chat overview: tick the relevant chats and click the icon next to the number of selected chats to move them together. Since late 2025, you can also create a new project from the selected chats in the same flow.
Since 10 July 2026, Claude’s memory has worked with separate, categorised memory items instead of a daily summary. Since 25 August 2026, everything Claude remembers is listed under Topics, where you can edit or delete each item, and memory is available on all plans, including Free. If a chat ends up in the wrong project by mistake, take it out with ‘Remove from project’.
Claude Projects vs Skills, Cowork and Artifacts: what is the difference?
A Claude project is not the same as a Claude skill, a Cowork session or a Claude artifact; this is a common source of confusion. They are four separate features that exist side by side, with different uses.
Projects are workspaces within claude.ai for grouping chats, instructions and files around one topic. Skills are reusable, structured instruction sets (in a SKILL.md format) that Claude calls up automatically when a particular task comes along, regardless of which project you are working in. Cowork is an agent that carries out real work (reading and creating files, running code) and, since 7 July 2026, runs on the web and mobile as well as desktop, where it is still in beta.
The practical dividing line: use a Project when you want to set the context for a series of chats on one topic. Use a Skill when you want to set out a specific way of working that applies across several topics. Cowork, which since 16 September 2026 has been gradually merging with the regular chat for Pro and Max, is meant for people who need agentic task execution, not just text output. And when Claude delivers something within a project that appears in its own window (a document, a dashboard, a mini app), you are dealing with artifacts: an output format that can exist alongside these three features and be used in any of them.
What are the most common mistakes with Claude projects?
The biggest mistake with Claude projects is not setting any instructions. A project without instructions is in effect a folder structure for chats and nothing more. The gain lies in the way instructions steer behaviour, not in grouping files.
A second common mistake is uploading every file you have on a topic, without selecting. Fifty PDFs in the project knowledge make retrieval worse, not better, because Claude has to filter out more noise. It is better to upload four or five well-chosen documents that cover the core and leave the rest out.
The third mistake is about maintenance. A project is not a static object. When your way of working changes, the instructions have to change with it. In my experience, projects slowly drift out of step with reality if you do not review them periodically. A simple routine that works: every two months, reread and adjust the instructions, and clear out files in the project knowledge that you no longer consult or that are out of date.
A fourth mistake is making the scope too broad. A project called ‘Work’ containing marketing, finance, HR and customer communication gives worse output than four separate projects, each with its own focus. The rule of thumb: one project per type of work, not one project per topic that happens to be on your mind.
A fifth pitfall I mainly see in organisations that are starting out with Claude: one person sets up a project and then does not share it in a structured way. As a result, the way of working stays in that one person’s head instead of being uniform across the department. For teams on a Team or Enterprise plan, sharing from day one is the rule, not the exception. If you want your team to work with Claude in a consistent, uniform way, the LearnLLM Claude Course teaches you how to build projects that keep working consistently across different roles and tasks.
Frequently asked questions about Claude Projects
Once a Claude project is set up, questions about connections and memory often follow.
Can you connect Google Docs to a Claude project?
Yes, you can connect Google Docs to a Claude project through the Google Drive connector, but only in private projects. Open the project’s Files section, click the plus sign and choose Drive, then search for a document or paste its URL. According to Anthropic, connected Google Docs sync directly from Drive, so you work with the latest version. Claude cannot see images in a Google Doc.
Does each Claude project have its own memory?
Yes, each Claude project has its own memory: what Claude remembers within a project is stored in that project. In a project, Claude uses that project memory together with the memory from your chats outside projects, but not the memory of other projects. If you want to keep a project completely separate, switch off the Use account memory option under Memory on the project page.
How do you delete a Claude project?
You delete a Claude project through the three dots next to the project, where you choose Delete and confirm with Yes, delete. You cannot delete an archived project; you have to unarchive it first. Archiving only hides a project from view: it stays accessible at the bottom of your project list and keeps its sharing permissions and members.
