Context management

Understanding Claude's Context Window

Context is Claude's working memory. Every file it reads, every command it runs, and every message you send takes up space in the context window. Think of the context window as the total amount of information Claude can hold in its memory at one time.

Whenever you enter a prompt, Claude reads a file, runs a tool call, or receives a tool call result, that data is added to the context window. Since there is only a finite amount of space available, it becomes extremely important to optimize this as much as possible.

Context Compaction and Clearing

When you approach the limit, the context window is automatically compacted. Compaction summarizes important details and removes unnecessary tool call results to free up space. However, note that this process could potentially lose some specific details from your previous conversation.

Manual Context Management

• /compact: You can run compaction manually with the /compact command. This will compact everything you have done up to that point. This is handy if you want to clear up context space while still retaining a memory of what you previously worked on.
• /clear: If you want to start completely from scratch without any memory of previous work, you can run the /clear command. This removes everything and resets the session.
• /context: To check the state of your memory, run the /context command. This provides a "big picture" view of your context size, showing which categories are consuming the most space, accompanied by a visual graphic.

Best Practices for Context Optimization

A general rule of thumb for managing your session:

• Compact when you are working on a specific feature and are hitting the context limit but need to continue. Keeping the context relevant to the current feature is important for ongoing development.
• Clear when you have finished a plan and want to start on a completely new feature. You do not want the previous conversation to present bias in the new tasks you want to create.

Persistent Memory with claude.md

For information that you want Claude to remember across different sessions, put it in the claude.md file. This ensures Claude doesn't have to rediscover core project details from scratch every time.

Prompt Specificity

The irony behind writing a smaller, vague prompt is that it can take up more context in the long run. Without being explicit, Claude is forced to look around your codebase more and perform its own "thinking," which consumes significantly more context window space than if you were just a little more clear with a sentence or two.

Advanced Optimization: MCP and Sub-agents

MCP Servers and Skills

MCP servers load all available tools into the context by default. If you have many MCP servers active for things unrelated to your current project, it might be worth turning them off. You can also try out Skills, which work similarly to MCP servers but do not put the entire toolset into context, saving you valuable space.

Sub-agents

Sub-agents run in parallel with your main agent but have a completely separate context window. For tasks that require an answer without needing the full "journey"—such as asking "Where are the authentication endpoints located?"—you can have a sub-agent do the work. It will return just a summary to your main agent, keeping the main context window clean.

Summary

Managing context within Claude Code is crucial for efficiency.

1. Use /compact to summarize long sessions.
2. Use /clear to start fresh and avoid bias.
3. Be specific with your prompts to reduce unnecessary codebase exploration.
4. Monitor your usage with /context.
5. Use sub-agents to delegate tasks where you only need the final answer.