Claude Code's Context Window, Explained

What's in there before you type a word

A common surprise: the window is already partly full at startup. Before your first prompt, Claude Code loads your CLAUDE.md, auto memory, MCP tool names, and your skill descriptions (not the skills themselves — full skill content loads only when one is actually used).

Then, as Claude works, each file it reads adds to the window, path-scoped rules load alongside the files they match, and hook output gets added too.

This is why "my context filled up fast and I barely said anything" is normal. Most of a typical session's context is project knowledge and tool results, not your words.

What happens when it fills up

A full context window does not end your session. Claude Code compacts automatically as you approach the limit — it summarises the conversation history so far and carries on.

The point at which that automatic pass runs depends on your model and configuration, so there is no single number worth quoting here. /context shows you where you actually stand, and the official docs list the per-model thresholds.

What survives compaction — and what doesn't

This is the part that trips people up, because it isn't uniform. Content loaded from disk comes back; content that lived in the conversation gets summarised away.

Two practical consequences worth knowing:

• A file over 5,000 tokens comes back as a path reference, not its content — it shows as Referenced file rather than Read. Its rules still reload.
• If a rule must survive compaction, don't scope it by path. Drop the paths: frontmatter or move it into the project-root CLAUDE.md.

Because truncation keeps the start of a skill file, put the instructions that matter most near the top of your SKILL.md.

Taking control before the automatic pass

Automatic compaction guesses what's important. You'll usually get a better result acting first:

• Compact with a focus. /compact focus on the auth bug fix keeps what you name instead of what the automatic pass infers. Worth doing before starting a long new task.
• Clear between unrelated tasks. /clear — old conversation crowds out the files you need next and costs tokens on every single message until it's gone.
• Compact earlier. /autocompact 500k sets how full the window gets before the automatic pass fires.
• Summarise part of a session. /rewind, pick a message, then Summarize from here or Summarize up to here.

The highest-leverage habit is delegating big reads to a subagent. A subagent works in its own separate context window, so the files it opens never touch yours — only its summary and a small metadata trailer come back. For research across many files, that's the difference between spending your window and spending someone else's.

Getting a bigger window instead of a smaller conversation

Sometimes the answer isn't a tighter session, it's more room. Fable 5, Sonnet 5, Opus 4.6 and later, and Sonnet 4.6 support a 1 million token context window.

Selection isn't identical across models: some are chosen as a [1m] model variant, while Sonnet 5 runs at 1M with no [1m] variant to pick. Availability also depends on your plan, so check the extended-context docs for your setup. Compaction still works the same way at the larger limit.

Be honest with yourself about whether you need it. A 1M window costs more per message once it's full, and a disciplined 200K session with subagents doing the heavy reading often beats a sprawling one.

Common questions

Related

• Claude's context window explained
• Claude Code commands reference
• Claude Code vs Cursor
• Claude MCP servers explained