Claude Code new sessions start with a lot of context, usually not because you just talked too much, but because you started loading too many project memories, long 'CLAUDE.md', large documentation, or MCP tool descriptions. Check '/context' first, and then decide what to delete.
Use /context to see the source first
The Anthropic documentation mentions '/context' to see what is currently loaded into the context. You need to know who is occupying the context: project files, chat history, tools, MCPs, memory files, or a bunch of rules documents.
Don't delete files based on feelings. Look at the proportion first, and then deal with the largest head.
CLAUDE.md Be short and precise
'CLAUDE.md' should write about project structure, common commands, coding rules, and key considerations, and is not suitable for long background materials, full API documentation, or historical decision journals.
If it is more than a few thousand words, it is recommended to split: leave only the entrance and index of the main file, put the detailed document in 'docs/', and let Claude Code read it when needed.
MCP tools also eat context
Every time an MCP server is connected, the tool name, description, and parameters may enter the context. With too many tools, Claude Code becomes slower, more expensive, and more likely to hesitate in tool selection.
Only the MCP required for the current mission is retained. For example, if you fix the front-end style, you don't need to use databases, browser automation, cloud deployment, and issue management all online.
Resolution Steps
First '/context', then shorten 'CLAUDE.md', then close extraneous MCP, and finally clean up irrelevant large documents. After doing this, open a new session to verify, and do not try repeatedly in an already contaminated session.
Conclusion: Context full is not a simple "model is too small", but more of a project loading strategy problem. Make the entry document shorter, fewer tools, and more focused tasks, and Claude Code will be stable.
