Understanding Total Context Window vs Usable Context Window in Claude Code

Introduction

One of the biggest misconceptions about Claude Code is that developers assume the entire 200k token context window is available for their prompts and project files.  In reality, only a portion of that context window is actually available for productive work.  The rest is already occupied by Claude Code's internal components.

The 200K Token Window

Claude Code operates within a 200,000 token context window.  Think of this context window as a fixed size container.  Everything Claude needs to perform a task must fit inside this container.  However, your prompts and project files are not the only things inside it.

What Occupies the Context Window?

Before you even type your first prompt, part of the context window is already reserved.

1. System Prompt (~6K Tokens)

This contains Claude's core instructions.

It defines:

• Safety rules
• Agent behavior
• Tool usage guidelines
• Response formatting

These tokens are fixed and cannot be used by the developer.

2. Tool Schemas (~8K Tokens)

Claude Code has access to tools such as:

• File reading
• File editing
• Bash commands
• Search tools
• MCP integrations

To use these tools, Claude must keep their definitions inside the context window.

These tokens are permanently reserved.

3. CLAUDE.md

Claude Code loads instructions from the project's CLAUDE.md file.

Although this file is usually small, it remains persistent throughout the session.

This means it continuously consumes a small portion of the available context.

4. Skills and MCP

When Claude Code starts, it loads:

• Skills
• MCP servers
• Tool integrations

These also occupy context space before any coding begins.

The Biggest Context Consumers

After startup, two things consume context faster than anything else:

Conversation History

In long sessions, conversation history often consumes 40–50% of the entire context window.

Every prompt and response remains in memory.  Surprisingly, Claude's own responses are often the largest contributor to context growth.

example:

You ask 10 questions
Claude writes 10 detailed answers

The answers themselves can occupy tens of thousands of tokens.

Tool Results

Tool outputs are another major source of context usage.

Examples:

• Reading files
• Bash command outputs
• Search results
• Grep results
• Logs

Imagine Claude reads:

auth.py (500 lines)
database.py (700 lines)
api.py (800 lines)

All of that content enters the context window.  This is why tool results are often called the "hidden context killer."

The Auto-Compact Buffer

One of the most interesting parts of Claude Code's architecture is the Auto-Compact Buffer.  Claude reserves approximately:

~33,000 tokens

for automatic compaction.  This space is not usable by the developer.  Its purpose is to give Claude room to summarize older conversation history before the context window becomes completely full.  Think of it as emergency memory reserved by the system. Without this buffer, long sessions would hit context limits much faster.

So How Much Context Is Actually Usable?

Many users hear:

200K Context Window

and assume they can use all 200,000 tokens.

In practice:

Total Context Window = 200K Tokens

- System Prompt
- Tool Schemas
- CLAUDE.md
- Skills + MCP
- Auto-Compact Buffer (~33K)

= Reduced Working Space

After accounting for reserved space, the practical working area available for:

• Prompts
• Project files
• Conversation history
• Tool outputs

is significantly smaller than the advertised 200K tokens.

Why Long Claude Code Sessions Slow Down

As a session grows:

Conversation History ↑
Tool Results ↑
File Reads ↑

The free space inside the context window decreases.

Eventually Claude must:

• Compress older information
• Summarize previous conversations
• Discard less relevant context

This process is called Auto-Compaction.

Once compaction begins, Claude starts replacing detailed history with summaries to make room for new information.