How to use Claude Projects, and make sure Claude reads what you gave it

Every month, an FP&A analyst opens a new chat and pastes the same three things: last month's board pack, the chart of accounts, and a paragraph explaining that "revenue" means recognised revenue, not bookings. Then she asks her actual question.

A project removes that ritual. Anthropic describes projects as "self-contained workspaces with their own chat histories and knowledge bases" (Anthropic Help Centre). Each one holds three things:

• Knowledge: files you upload once, which Claude uses as background for every chat in the project.
• Instructions: standing guidance on how to answer, applied to every chat.
• Chats: the project's own conversation history, kept apart from your other chats.

Projects also have their own memory. Inside a project, Claude uses that project's memory alongside your general memory, and never another project's (more in Claude memory).

Projects launched in June 2024 for Pro and Team with a 200,000-token context, which Anthropic called "the equivalent of a 500-page book" (Anthropic, June 2024). Claude's plans now list a context window of "up to 1M", varying by model (Anthropic pricing, October 2026). That is a lot of board packs. Whether Claude reads all of them carefully is a separate question, covered below.

Most weak projects have plenty of files and almost no instructions. Reverse that.

1. Create the project and give it a name that says what it's for and when: "Board pack FY27" beats "Finance".
2. Write the instructions. Say who the work is for, define the terms that matter, and say what to do when the files don't cover a question. Anthropic's advice for reducing hallucinations is to give Claude explicit permission to say it doesn't know (Anthropic, Reduce hallucinations).
3. Upload the fewest files that do the job. Each file can be up to 30 MB. There's no fixed count, but together they must fit the context window. Only PDFs get read as images as well as text; for other files Claude extracts the text.
4. Date your files. Put the period in the file name ("pricing-2026-Q3.pdf"). When something is superseded, delete the old version rather than adding the new one beside it.
5. Share it, on Team or Enterprise. Give colleagues "Can view" or "Can edit" by email, in bulk, or across the organisation. Owners can switch sharing off.

Instructions that work for a finance project (illustrative):

You're helping the FP&A team prepare monthly board materials for the CFO.
Definitions: "revenue" means recognised revenue per the policy doc in this
project. "ARR" excludes one-off services. Fiscal year starts 1 February.
Use figures only from the files in this project or data I paste in.
Quote the file name and section for every figure.
If the files don't answer the question, say "Not in the project files"
and stop. Don't estimate.

The last two lines matter most. Without them, a gap in the files becomes a plausible number.

Those lines lower the odds of a made-up number. They don't remove it. With Count connected, every answer in the project shows which claims rest on your files and which rest on nothing.

Anthropic draws the line clearly: projects hold "always-loaded" knowledge, skills are loaded only when a task needs them, and connectors (MCP) link Claude to live services (Anthropic Help Centre, What are skills). Memory adds facts Claude picked up along the way. In practice:

The common mistake is treating project knowledge as a database. A CSV exported last Tuesday is a snapshot that ages from the moment you upload it. If the number needs to be today's, connect the source instead.

Projects are best for long-running work where the background changes slowly: a due-diligence review, a research programme, the monthly close. Everyone on the team asks from the same starting point, which removes a whole class of "it gave me a different answer" arguments.

They are weakest in three situations.

When files age. Nothing tells Claude a file is out of date. Upload this year's price list without removing last year's and you have two price lists, and Claude may use either.

When the knowledge base gets big. On paid plans, when project knowledge nears the context limit, Claude switches to RAG mode, which expands capacity by up to 10x (Anthropic Help Centre). RAG means retrieving the passages that look most relevant instead of reading everything each time. That's how it fits more in. It also means a relevant passage can be missed when the question is worded differently from the document.

When the context is simply long. Anthropic's engineering team describes "context rot": recall of any one item falls as the context grows (Anthropic Engineering, September 2025). A project stuffed with 40 files gives Claude more to lose track of, not more to work with.

A new version of projects is also in beta, rolling out in stages from Claude Code: each project is one conversation, and Claude splits work into parallel threads that run in the cloud. Existing projects moved to it keep their instructions, files, history and scheduled tasks; running several threads at once uses your plan's allowance faster.

The most common project failure isn't a hallucination from nowhere. It's an answer that ignores, or misreads, a file you uploaded.

The research on long documents is sobering. In FinanceBench, even when GPT-4-Turbo was handed the correct page of a company filing, it got 15% of questions wrong (Patronus AI, FinanceBench, November 2023). That's an older model from a different vendor, and current Claude models are much stronger, but the lesson holds: having the right document in context doesn't guarantee the right reading of it. Anthropic's own guidance for long material is to ask Claude to quote the relevant passages before answering, and to put the question after the documents (Anthropic, Prompting best practices).

A two-minute check on a project answer

1. Ask for the quote: "Which file and which section is that from? Quote the sentence."
2. Find the quote yourself in the file. Ctrl+F is enough. If the words aren't there, the answer isn't grounded in the file, whatever Claude says.
3. Plant a canary. Ask a question whose answer you know sits deep in your largest or oldest file. If Claude misses it, it's missing other things too.
4. Check for duplicates. Does the project hold two versions of anything? If so, ask which one the answer used.
5. Ask about the instructions: "Did any part of my project instructions conflict with this answer?" Claude will often point to the rule it bent.

Steps 1 and 2 are the manual version of a citation check. Count does it for a whole answer: connect it to Claude and every claim is labelled as a citation you can open, an assumption, model knowledge or a hallucination, with the reasoning chain under it.

How do I use Claude Projects?

Are Claude Projects free?

How many files can I add to a Claude project?

Do chats in a Claude project share memory?