- Tool
- Claude Projects
- Price
- Included, free up to 5 projects
- Time
- 30 min
- Level
- Beginner
A customer asks about the cancellation period, an employee asks how product returns are handled. The answer exists, somewhere in a PDF. You open three files, search, and rephrase. A general-purpose AI, meanwhile, may answer with what it “knows” from the internet rather than with your own rules. The goal here: an assistant that only talks about your documents.
First: why not NotebookLM plugged into Claude
NotebookLM (now Gemini Notebook at Google) already does this job very well on its own. Plugging it into Claude is trickier: the consumer version has no official interface for programs. The most followed open-source project, notebooklm-mcp-cli (MIT licence), says so in its own disclaimer: it uses undocumented internal functions, extracts the cookies of your Google account and may not be aligned with Google's terms. Its author reserves it for personal, experimental use. For business documents, I do not recommend it. The serious option is to do the job directly in Claude, with a Project.
What a Claude Project is
A Project is a workspace with its own chat history, its own knowledge base (your files) and its own instructions. Everything you add serves as context for every chat in the project. Paid plans add document search (RAG): when there are too many files to read at once, Claude fetches only the useful passages, which raises capacity by up to ten times according to Anthropic. On Team and Enterprise plans, a project can be shared with colleagues.
Create the project step by step
- Prepare the documentsGather the current versions, one per topic, with clear names (
terms-2026.pdf,prices-2026.pdf). A scanned or messy PDF reads poorly: convert it to text first, as explained in the tutorial Convert PDF and Word to Markdown for AI. - Create the projectOn claude.ai, open Projects, create one and name it (“Customer base”, “Internal procedures”).
- Add the filesUpload them to the project's knowledge section.
- Write the instructionsPaste this into the project instructions:
You answer questions about [company name] using only the documents in this project. Do not use your general knowledge. For each answer: 1. First find the passages in the documents that answer the question. 2. Answer in plain English, in 2 to 5 short sentences. 3. End with the source: document name and an exact quote of the passage, in quotation marks. If the documents do not contain the answer, say exactly: "I can't find this information in your documents." then say which document should be added. Never make up a price, a deadline or a rule. If two documents contradict each other, quote both and flag it. - Test with questions you know the answer toAsk five verifiable questions, including one that is in no document. If Claude invents instead of saying “I can't find…”, tighten the instruction.
Example in a small business
An online shop puts its terms of sale, returns policy, shipping rates and template replies in a project. When a customer message comes in, the team asks the project:
A customer received a damaged item 20 days ago and is asking
for a refund. What do our documents say? Then draft a short,
polite reply that quotes the rule that applies.Claude answers with the exact rule and the quote, then suggests a draft. A person reviews it before sending.
Limits and cost
- These instructions greatly reduce made-up answers without eliminating them: Anthropic says so in its documentation. Check the quote before any answer that commits you (price, deadline, warranty).
- The project does not update itself: replace the file when your terms change, and delete the old version to avoid contradictions.
- Free plan: five projects at most and no extended search (RAG). For a large document base, you need a paid plan.
Ground rules. Do not put passwords, bank details or named customer data in a project. For confidential documents, check your account's privacy settings and prefer a business plan (Team or Enterprise). If you still want to try an unofficial NotebookLM bridge, do it with a separate Google account, never the company one.
Sources
- Claude Help Center: What are projects
- Claude Help Center: RAG for projects
- Anthropic documentation: Reduce hallucinations
- notebooklm-mcp-cli on GitHub (unofficial project)
Want the same assistant to answer your customers directly, on your website or by email, with a human check? That is what I build as AI agents.