Claude Code's context window has been upgraded to 1 million tokens, but the more space you have, the harder it is to manage. This article comes from the hands-on experience of Thariq, an official Anthropic evangelist, and it tackles the same problem: how to manage your context so Claude Code actually helps you get work done instead of making a bigger mess.

Who this is for: developers already using Claude Code, especially those who frequently hit “Claude forgot what we said earlier” or “quality drops sharply after compaction”.

Core Concepts: Three Things About the Context Window

Context window: everything the model can “see” when generating a response, including the system prompt, chat history, tool calls and their outputs, and files it has read. Claude Code currently has 1 million tokens of context capacity.

Context rot: the longer the conversation history, the more scattered the model's attention, and important early information can get drowned out. This is an unavoidable physical limit.

Compaction: summarizing a long conversation into a concise digest to free up space and keep working. This is a “lossy” operation — you hand the judgment of “what matters” over to Claude.

Five Options After Each Turn

After Claude finishes a response, you have five paths:

Tip 1: Rewind Instead of Correcting

This is the single most important habit recommended by Anthropic. When Claude tries an approach and it fails, your instinct is to tell it “switch to approach X”. But the better move is:

1. Double-tap Esc to rewind to the point right after it finished reading the file
2. Re-prompt with the lesson you just learned: “Don't use approach A anymore — the foo module doesn't support that. Try approach B directly”

The benefit: the noise of the failed attempt is completely removed from the context instead of continuing to interfere in later turns.

Tip: You can also use the “summarize from here” feature to have Claude summarize the lessons from its stumbles as a handoff note — like “future Claude” leaving a note for its past self.

Tip 2: Compact or Clear — Which One?

Compact (/compact): Claude summarizes the conversation history itself and replaces the original with the summary. Less effort, but you don't fully control what it keeps.

You can add instructions to steer the compaction:

/compact Focus on the refactoring of the authentication module and drop the test-debugging content

Clear (/clear): you write down the key points yourself and restart from a clean state. More effort, but the new context is 100% what you consider important.

Rule of thumb: compact within a task, clear when switching tasks.

Tip 3: When Should You Start a New Session?

The basic principle: new task, new session.

1 million tokens let you complete longer, more complex tasks (like building a full-stack application from scratch), but tightly linked follow-up tasks are the exception. Say you just finished a new feature and now need to write its documentation — starting a new session is actually slower, because Claude has to read through all the code again.

Tip 4: Use Subagents for Read-Once-and-Done Work

Subagents have their own context windows and return only their results to the main session when the work is done. The test for whether to use one: will you ever need to look at these intermediate results again? If not, send a subagent to do it.

Real-world usage examples:

• “Send a subagent to verify our recent work against this spec document”
• “Send a subagent to read through another codebase, summarize how it implements authentication, then mirror that implementation over here”
• “Send a subagent to write documentation for the new feature based on the Git change history”

Why Does Compaction Go Wrong?

Compaction failures usually happen at one specific moment: when the LLM cannot predict what you will do next. For example, auto-compaction triggers after a long debugging stretch, and then you say “also fix that other warning we saw earlier in bar.ts” — but that warning was already discarded as irrelevant during compaction.

The countermeasure: with 1 million tokens of headroom, you can proactively run /compact ahead of time and include a description of “what I plan to do next”.