Context Window and How Projects Extend It
Understanding Claude's context window is essential to using Projects effectively. This lesson explains what the context window is, why it matters, and how Projects help you work around its limits.
What Is the Context Window?
The context window is the total amount of text Claude can "see" at once during a conversation. Think of it as Claude's working memory — everything it can consider when generating a response.
| Model | Context Window | Approximate Word Count |
|---|---|---|
| Claude Sonnet | 200K tokens | ~150,000 words |
| Claude Opus | 200K tokens | ~150,000 words |
| Claude Haiku | 200K tokens | ~150,000 words |
A 200K token context window can hold roughly:
- A 500-page book
- An entire codebase of a medium-sized application
- Hours of meeting transcripts
- Dozens of research papers
This is significantly larger than many competing models, but it is still finite.
How the Context Window Gets Used
The context window is shared between several things:
Context Window (200K tokens)
├── System prompt and instructions
├── Project custom instructions
├── Project knowledge files (loaded on demand)
├── Conversation history (all messages so far)
├── Current user message
└── Claude's response
As your conversation grows longer, earlier messages consume more of the window. Eventually, older context may be summarized or dropped to make room for new messages.
The Long Conversation Problem
In a regular (non-Project) conversation, this creates a practical limit:
- You start a conversation explaining your context, preferences, and goals
- You have a productive discussion for 10-20 messages
- The conversation gets long and Claude starts losing track of earlier details
- You start a new conversation and have to re-explain everything
This cycle wastes time and loses accumulated context.
How Projects Solve This
Projects address the long conversation problem in two ways:
1. Custom Instructions Are Always Present
Your custom instructions are loaded at the start of every conversation in the project. They do not count against your conversation history — they are always available:
- Claude always knows your role, preferences, and constraints
- You never need to re-explain "I'm building a Next.js app with TypeScript"
- Your formatting preferences and rules are applied automatically
2. Knowledge Files Extend Effective Memory
Knowledge files act as a persistent reference library:
- Without Projects: You upload a PDF, discuss it, and it disappears when you start a new conversation
- With Projects: Your files are always available across every conversation in the project
This means Claude can reference your brand guidelines in conversation 1, your API documentation in conversation 2, and your competitor analysis in conversation 3 — all without re-uploading.
Managing Context Effectively
Even with Projects, you should be strategic about how you use the context window.
Start New Conversations for New Topics
Within a project, start a new conversation when you shift to a different topic:
| Instead Of | Do This |
|---|---|
| One 50-message conversation covering design, backend, and testing | Three separate conversations: "Homepage Design," "API Endpoints," "Test Strategy" |
Each new conversation starts fresh but still has access to your project instructions and knowledge files.
Summarize Before Context Gets Long
If a conversation is getting long and productive, ask Claude to summarize:
Summarize the key decisions we have made so far in this conversation.
I want to save this before we continue.
Copy this summary and paste it at the start of a new conversation if needed.
Prioritize Your Knowledge Files
Projects have a total knowledge base size limit. Prioritize the most important files:
High priority (upload first):
- Configuration files and schemas
- Style guides and brand documents
- Core reference materials
Lower priority (summarize instead):
- Very long research papers (upload an executive summary instead)
- Entire codebases (upload key files: config, schema, main modules)
- Historical data (upload a summary with key metrics)
Use Explicit References
When you need Claude to focus on a specific uploaded file, reference it directly:
Using the API schema I uploaded (api-schema-v2.json),
generate TypeScript types for the User and Order endpoints.
This is more reliable than hoping Claude will find the right file on its own.
Context Window vs. Knowledge Files: What Goes Where
| Information Type | Where It Goes | Why |
|---|---|---|
| Preferences and rules | Custom instructions | Always active, does not clutter files |
| Reference documents | Knowledge files | Persistent across conversations |
| Current task details | Conversation message | Specific to this chat |
| Previous decisions | New conversation starter | Carries forward context |
Practical Example: Research Project
Here is how context management works in practice for a research project:
Project setup:
- Custom instructions: "You are helping me write a literature review on AI in healthcare. Use APA citation format. Be critical of methodology in the papers I share."
- Knowledge files: 5 key research papers (PDFs), your thesis outline (Markdown)
Conversation 1: "Summarize the methodology of each uploaded paper"
- Claude reads all 5 papers and creates a comparison table
Conversation 2: "Based on the papers, identify gaps in current research"
- Claude references the same papers without re-uploading
- Builds on the context from instructions
Conversation 3: "Draft the literature review introduction"
- Uses your thesis outline, the papers, and the writing guidelines from instructions
- No re-explanation needed
Each conversation is focused, manageable, and builds on the shared project context.
Key Takeaways
- The context window (200K tokens) is Claude's working memory — shared between instructions, files, conversation, and responses
- Long conversations eventually lose earlier context as the window fills up
- Projects solve this by keeping custom instructions and knowledge files persistent across conversations
- Start new conversations for new topics — they still access project instructions and files
- Prioritize knowledge files by importance and summarize very long documents before uploading
- Reference specific files by name when you need Claude to focus on particular material