Mastering Claude Code: Session Management Strategies for 1M Context Windows

Why a Bigger Window Isn’t Enough

Anthropic’s recent blog post shows that a 1 M‑token window lets Claude Code handle longer tasks, but merely increasing size does not guarantee success. As the context grows, attention becomes diluted and context rot appears, typically around 300 k–400 k tokens, causing older, irrelevant information to hinder the current task.

What Lives Inside the Context Window

The window stores more than the dialogue history. It includes the system prompt, the CLAUDE.md configuration file, the full conversation, every tool call with its result, and every file read during execution. This makes the window a dynamic working memory that accumulates state as the task progresses.

Five Decision Points After Each Model Output

When Claude finishes a turn and you are about to send the next message, you actually have several choices.

Continue : keep all history and proceed.

/rewind (double‑tap Esc): jump back to a previous message, discarding the failed tail while preserving useful prefixes.

/clear : start a brand‑new session with a hand‑written brief that defines the new constraints.

Compact : ask the model to summarize the current session and replace the long history with a concise abstract.

Subagent : spawn an isolated context to perform noisy intermediate work and return only the final conclusion.

These actions map to classic software‑engineering concepts: retaining, discarding, or isolating state.

When to Use Each Action

If the current task still benefits from the existing context, Continue is fine.

If the model has taken a wrong path, use /rewind to cut off the erroneous branch.

If the session is bloated with exploration noise, Compact can compress it—preferably with a prompt that tells the model what to keep.

When starting a truly new task, /clear lets you rebuild a clean context.

For operations that generate a lot of intermediate output (file reads, tool calls, dead‑ends), delegate to a Subagent so the main thread only receives the final report.

Compact vs. /clear

Compact performs a lossy, model‑driven summarization; the model decides what stays, which can be risky when attention is already scattered. /clear is a manual, lossless reset where you explicitly write the brief, ensuring only verified information remains.

Subagent Isolation

A Subagent runs in its own context, keeping all intermediate noise (failed attempts, file reads, tool outputs) isolated. The main thread receives only the final report, preventing the main working memory from being polluted.

Practical Cheat Sheet

Continue : use when the existing context is still helpful.

/rewind (double‑Esc): revert to the point before a mistake, preserving useful prefixes.

/compact <prompt> : compress a long session, specifying what to retain.

/clear : write a fresh brief and start a new session.

Subagent : employ when the next steps will generate massive intermediate noise.

Takeaway

Long‑running AI tasks are limited not just by window size but by how well the working memory is managed. Choosing the right action—Continue, /rewind, Compact, /clear, or Subagent—can keep the session focused, reduce context rot, and improve overall reliability.