Claude Projects are by far the most powerful feature Claude offers when you work with the same documents repeatedly. A dedicated knowledge base that is automatically available in every chat within the project. But the quality of your answers depends directly on how you build that knowledge base. Poorly structured documents lead to poor answers. This guide shows you how to prepare documents so that Claude actually finds and uses them correctly.
How Claude Projects work
A Claude Project is essentially a container where you can store three things: System Instructions (the directives that apply in every chat), Project Knowledge (your document base), and chat history. In every new chat within the project, Claude automatically has access to System Instructions and Knowledge.
The key point: Claude does not read all knowledge documents in every response. For small projects, everything fits into the context window and Claude uses it directly. For larger knowledge bases, Claude performs a kind of internal retrieval, searching for relevant passages in your documents and using only those. How well this internal search works depends entirely on how your documents are structured.
The context window of Claude Opus and Sonnet is 200,000 tokens. That corresponds to about 150,000 words or 350 pages of text. Anthropic offers a 1-million-token window in beta for enterprise customers.
What goes wrong with unstructured documents
Most users just upload what they have: PDF reports, Word documents, sometimes Excel sheets, occasionally PowerPoint. This works okay in 80 percent of cases and really badly in 20 percent. The typical problems:
PDFs with multi-column layouts. Scientific papers, magazines, reports. Claude often reads them in the wrong order, jumps between columns, mixes content from different sections.
Excel tables with complex structures. Nested headers, merged cells, pivot tables. Claude often does not understand the relationships and responds with incorrect numbers.
PowerPoint slides without semantic structure. When content sits only in text boxes instead of slide titles and bullet lists, the hierarchy is lost.
Images without text recognition. Scanned PDFs without OCR are worthless to Claude. He sees only the image, not the content.
Single documents that are too long. A 500-page PDF in a project is usable, but the internal search becomes less precise.
Three rules for good project knowledge
From practice, three simple rules have proven to dramatically improve response quality in most setups.
Rule one, Markdown above everything. If you have PDFs, Word documents, or Excel sheets, convert them to Markdown before upload. You can use MarkItDown in your browser, which handles the conversion locally without your documents ever reaching a server. Markdown gives Claude a clear hierarchy of headings and sections that it can understand semantically.
Rule two, one document per topic. Better ten clearly separated Markdown files than one 500-page collection. Claude finds relevant passages faster when each document has a clearly defined topic. Name files meaningfully: product-handbook-v3.md instead of document1.md.
Rule three, metadata at the document start. Begin each Markdown file with a short block that tells Claude what the file contains:
# CRM Product Handbook v3.2
**Updated:** May 2026
**Content:** Complete user manual for the CRM software, including onboarding, features, troubleshooting.
**Audience:** End users, support team.
This metadata helps Claude during internal search to find the right document.
Practical example: setting up a consulting project
A concrete setup for a typical consulting project with a client:
ClaudeProject: Client XYZ Digital Transformation
System Instructions:
"You are a consultant for client XYZ. Respond factually, use the
knowledge documents for factual statements, acknowledge when something
is not contained in the documents."
Knowledge:
01-client-profile.md (Industry, size, stakeholders)
02-project-history.md (Previous milestones)
03-current-processes.md (As-is state of workflows)
04-target-state.md (Target architecture)
05-contracts.md (Master agreements, NDAs, SLAs)
06-meeting-notes.md (Minutes from all workshops)
07-technical-specs.md (Architecture, APIs, databases)
When you now ask in chat "What did we discuss about API authentication in the last workshop?", Claude searches specifically in 06-meeting-notes.md and 07-technical-specs.md. The answer becomes precise because the documents are clearly separated.
Use numbered prefixes in filenames (01-, 02-, 03-). This helps Claude with natural sorting and you with overview.
Understanding and using the token budget
The 200,000-token context window sounds huge, but it is not when you work systematically. A rough breakdown in practice:
| Area | Typical token usage |
|---|---|
| System Instructions | 200 to 1,500 |
| Knowledge documents (loaded per query) | 2,000 to 30,000 |
| Chat history | 500 to 50,000 |
| Current user question | 50 to 500 |
| Claude's response | 500 to 8,000 |
| Buffer for internal processing | approximately 10,000 |
If you have a project with 50 documents and several hours of chat history, the context window can become tight. Claude then starts to summarize older chat messages or use less knowledge.
Very long chat history in a project eats tokens that are missing for knowledge. When you notice responses becoming less precise, start a new chat in the same project. Knowledge and System Instructions remain, only the history resets.
Which file formats you should actually use
Claude accepts many formats in the knowledge area, but quality varies significantly.
Markdown (.md). The best choice for everything. Clear hierarchy, compact, semantically unambiguous. If you take away only one recommendation, take this.
Plain text (.txt). Second-best option. When you truly have no structure or the file was auto-generated.
PDF. Works, but not ideal. Claude has to extract text internally, and structures often get lost. Convert PDFs to Markdown first with MarkItDown.
Word and Excel. Supported, but Excel tables are often misinterpreted. Word is okay, but Markdown is cleaner. Conversion is worth it.
Images. Only useful when you want to reference diagrams or screenshots. Plain text in images is often not recognized, OCR is limited.
Code files (.py, .js, .ts, etc.). Work well, Claude understands the code directly. No conversion needed.
A concrete workflow to copy
When setting up a new project, you can follow this workflow:
- Collect all relevant source documents in one folder on your machine.
- Convert every file that is not already Markdown or text to Markdown. You can use MarkItDown in the browser, which covers PDF, Word, Excel, PowerPoint, images, and HTML.
- Open each converted file briefly and add a metadata block at the top (three to five lines: title, date, content, audience).
- Name files clearly and number them in logical order.
- Upload the files to your Claude Project.
- Write System Instructions that explicitly reference the knowledge documents ("Respond based on the stored documents, indicate when something is not contained").
- Test the setup with three to five typical questions your project should be able to answer.
When your test questions are answered poorly, the cause is almost always document structure, not Claude. Look into the knowledge documents and check whether the relevant information is clearly findable.
Common mistakes and how to avoid them
Mistake one, packing all knowledge into one file. A 200-page Markdown file is worse than twenty ten-page files. Claude finds relevant content faster when documents are topically focused.
Mistake two, uploading original files without conversion. A 50-MB PowerPoint file with 200 slides full of animations and images is largely useless to Claude. Convert to Markdown, drop the images, keep only the content substance.
Mistake three, no System Instructions. Without clear directives, Claude does not know how to use the knowledge. Minimum two sentences: what is the task, how should the documents be handled.
Mistake four, never updating knowledge. As projects evolve, the documents in knowledge should grow too. Review monthly whether everything is current.
Mistake five, uploading sensitive data without anonymization. Claude Knowledge is private, but before uploading customer data or confidential information, check the Anthropic data policy for your plan. Enterprise plans have additional guarantees, private plans have fewer.
Conclusion
Claude Projects become as good as your knowledge base is. Three simple rules make the biggest difference: Markdown instead of original formats, one topic per file, metadata at the document start. If you implement these three points consistently, you get noticeably more precise responses in every chat within the project. The one-time effort for preparation pays off after a few weeks of use.
Read more
- Use MarkItDown directly to convert your source documents to Markdown
- Convert PDF to Markdown for PDF-specific conversion
- Convert Word to Markdown for DOCX files
- What is Markdown? if you haven't read the basics guide yet