Anthropic explains when to start a new session in Claude Code, and how

A large context window in Claude Code gives developers more room to work, but it also raises the cost of mismanaging a session. The longer a single conversation runs, the more likely it is to pick up irrelevant files, stale decisions, and half-finished hypotheses, and that clutter can hurt the model as much as useful data helps it. That is why Anthropic added a /usage command to Claude Code and called out how developers run their sessions.

Once the window grew to 1 million tokens, it became clear that people work in very different ways: some keep a single long terminal conversation going for weeks, while others start a new session after almost every request. The difference isn't cosmetic. Your strategy affects the quality of the prompts, how the model behaves during compaction, and what the work ultimately costs.

The core idea is simple: there is no universal rule that says to keep one endless session or, conversely, to start over constantly. A long session pays off when you are working through one connected task, such as mapping out a project's architecture, editing several files, and gradually refining requirements. In that mode, accumulated context speeds things up: the model doesn't have to reread the codebase, recall constraints, or reconstruct decisions already made.

A new session makes sense in a different situation: when the task has fundamentally changed, the conversation has filled up with scratch-work debris, or the model has latched onto a wrong assumption. The signs usually show up quickly. Claude Code starts leaning more on old lines of reasoning, proposes fixes for problems already solved, circles back to abandoned ideas, or pulls irrelevant details from early messages into its answers.

At that point, the large context stops being an advantage and becomes a source of noise. Restarting the session clears out the excess baggage and lets you reframe the task. That explains the interest in context-management tools inside Claude Code.

The /usage command isn't just there to satisfy curiosity; it shows how much of the budget is already used and whether it's time to rebuild the conversation. The compact feature helps when a dialogue still covers a single task but needs to be condensed down to what matters. Rewind lets you roll back to an earlier point if the discussion has wandered off course.

Subagents are worth bringing in when part of the work can run in a separate thread without loading the main context with side investigations. In every case the goal is the same: keep only what actually helps with the next step in the window. The catch is that compaction isn't a magic button.

If a session is already packed with questionable decisions, duplicate instructions, and code fragments that no longer matter, that noise can survive compression in a more concentrated form. So the best results usually come not from trying to rescue an overloaded conversation at any cost, but from making the call in time: either trim it carefully around a single goal or simply start a new conversation from a clean slate. In essence, Anthropic is describing for Claude Code the same discipline experienced developers have long applied to code and tasks: separate concerns, don't hoard what you don't need, and regularly check that the system is still solving the current problem rather than fighting the legacy of earlier decisions.

A million tokens doesn't do away with that discipline; it makes it even more important. The more memory is available, the higher the cost of poor structure, because it holds not just useful context but far more confusion. The practical takeaway: keep a Claude Code session going as long as it helps you move forward on one clear task, and restart it as soon as the context starts working against the goal.

/usage now makes this easier to track, and compact, rewind, and subagents are no longer minor helpers but core quality tools. For developers, that means more accurate answers, fewer false starts, and predictable behavior even over very long stretches.