Claude Context Window Full: Practical Fixes

A full context window is a different problem from a usage limit, and the fix is to give Claude less to carry, not to buy more capacity.

What you are seeing

The context window is Claude's working memory: everything in the current conversation, plus project instructions, attached files and tool output. When it fills, Claude either stops accepting new messages in that chat, starts summarising earlier turns, or in Claude Code shows a message such as Prompt is too long or Context limit reached · /compact or /clear to continue.

It is not the same as "You've hit your session limit" or a weekly limit. Those are about how much you have used across all chats in a window of time. A context problem is about one conversation being too big. Upgrading a plan does not fix it, and the five-hour usage limit note covers the other case.

At the time of writing, Anthropic states that on paid plans the newest models support up to a 1M token context window, while others support 500K or 200K tokens. A 200K window is roughly a long novel, which sounds generous until a team starts pasting whole spreadsheets and month-long threads into one chat.

Likely causes, most common first

1. One chat used for everything. A single conversation that has run for weeks carries every earlier draft and every attached file on every turn.
2. Large project knowledge and long instructions. Files and instructions in a Project are loaded for every chat inside it.
3. Big attachments or tool output. A 300-page PDF, a full export, a whole log file, or a long web research run.
4. Extended thinking and high effort on routine tasks. These add tokens per turn on paid plans.
5. In Claude Code, an oversized file or command output. Reading a generated bundle or a 10,000-line test log fills the window in one go, and a single huge paste can leave /compact with nothing to summarise.

Fix in claude.ai

1. Start a new chat for the next task. Ask Claude to write a short handover first: the decisions made, open questions and the final version of any text. Paste that into the new chat.
2. If the material is genuinely large, put it in a Project rather than the chat. When project knowledge approaches the context limit on paid plans, Claude switches to retrieval mode and pulls in the relevant parts rather than everything.
3. Trim project instructions to what applies to every chat. Remove project files nobody has referenced in a month.
4. Turn off extended thinking and lower the effort level for routine drafting. Turn off connected apps and web search when the task does not need them.
5. If you see Claude "organising its thoughts" in a long chat, that is automatic context management working. It is a cue to wrap up, not a fault.

Our note on Claude Projects context management goes into how to structure a project so this rarely happens.

Fix in Claude Code

1. Run /context to see what is using the space: conversation, CLAUDE.md, MCP tools, files.
2. Run /compact with a focus, for example /compact keep only the plan and the diff. If it returns Not enough messages to compact, the window was filled by one large item, so go to step 4.
3. If the earlier conversation is no longer needed, /rename the session so you can find it later, then /clear.
4. Stop reading whole files. Ask for a line range or a single function. Send verbose work such as test runs or log analysis to a subagent so only a summary returns.
5. If you see Autocompact is thrashing, a file or output is refilling the window after every compaction. Read it in smaller chunks or move that work to a subagent.
6. Keep CLAUDE.md under about 200 lines and move workflow-specific instructions into skills, which load only when invoked. Disable MCP servers you are not using with /mcp.

What the admin should check

• Whether the team has a written habit of one chat per task and a handover note between chats. This removes most of the problem for non-technical staff.
• Shared Projects that have grown into dumping grounds. Assign an owner and a quarterly tidy-up.
• Claude Code repositories with large generated files not in .gitignore.
• The /usage breakdown on Pro, Max, Team and Enterprise plans, which flags long context when it accounts for a large share of recent usage. Long context also costs allowance, so a context habit is a budget habit.

When to contact Anthropic

Contact support when a new chat with a small prompt still reports a context or length error, or when Claude Code says the prompt is too long right after /clear. Use Get help from the lower left of claude.ai and include the plan, the model, the exact message and the time. Do not attach the confidential document that triggered it; describe its size instead.

Source basis

• Claude support: understanding Claude's context window
• Claude Code: troubleshooting
• Claude Code: manage costs effectively
• Claude Code: error reference