Mastering Claude Code’s 1M Context: Anthropic’s Five Essential Management Strategies

Claude Code’s context window determines how much input the model can consider at once. Before Opus 4.6 it was limited to 200 k tokens, but the upgrade to a 1 M token window introduces new challenges such as “context rot,” where attention drifts as the context grows, degrading performance.

0 Context Rot

Context rot describes the gradual decline in model quality when the context becomes long; early, irrelevant tokens start to distract the model. The problem becomes more noticeable with the 1 M context because more material can be packed in, increasing the pressure on the model to decide what is important.

1 Continue – The Simplest Action

When the current context still contains useful information for the next step, the recommended action is to simply continue. The official rule is: if the needed data is still present in the context, do not compact or clear, as rebuilding the context would waste tokens.

2 Rewind – Erase Failed Attempts

The /rewind command rolls back the session to any previous message, discarding everything after that point. For example, after Claude reads five files and a strategy fails, instead of appending a correction (“the method A doesn’t work, try B”), you can rewind to the moment after the file reads and issue a fresh prompt that directly states the new instruction, removing the failed attempt from the context.

Rewind can be paired with /summarize from here to let Claude generate a hand‑off note before rewinding, effectively leaving a reminder for future steps.

3 Clear – Start a Fresh Session

The guideline is to open a new session for a new task. Using /clear discards the existing context and lets you write a concise brief that includes only the essential information for the new task, avoiding the overhead of irrelevant previous context.

4 Compact – Let the Model Summarize

The /compact command asks the model to summarize earlier dialogue and replace the original history with that summary. This saves space but can lose details the model does not know you’ll need later. Manual compacting with a focused prompt (e.g., /compact focus on the auth refactor, drop the test debugging) can guide what to keep.

Automatic compacting (autocompact) can fail when it triggers while the model’s attention is already scattered, causing important but low‑frequency information to be dropped.

5 Subagents – Spawn a Helper When Only the Final Result Is Needed

Subagents should be used when you only need the final conclusion, not the intermediate outputs. For instance, searching a large codebase for a keyword and summarizing the findings generates many intermediate reads that are irrelevant if you only care about the final summary. In such cases, you can launch a subagent with prompts like:

Spin up a subagent to verify the result of this work based on the following spec file

Spin off a subagent to read through this other codebase and summarize how it implemented the auth flow, then implement it yourself in the same way

Spin off a subagent to write the docs on this feature based on my git changes

After Opus 4.7 the default behavior became more conservative, so you must explicitly request a subagent in the prompt.

Summary and Comparison

The article concludes with a comparison table (shown in the original images) that maps each of the five actions to typical scenarios, helping users decide which command to issue. It also mentions the new /usage slash command for inspecting Claude Code’s resource consumption, which aids in making cost‑effective session‑management decisions.

Overall, the expanded 1 M context provides more flexibility, but effective management of that context—through the five actions above—is essential to keep Claude Code performant and economical.