Context Window Management

The context window is Claude’s working memory (~200K tokens standard; Sonnet 5 supports a native 1M window, and opus[1m] selects 1M for Opus). Everything competes for it: system prompt, CLAUDE.md, auto memory, tool schemas, file reads, tool output, conversation. Managing it well is the #1 skill separating intermediate from pro users.

What Loads at Startup (before you type anything)

Item	Approx. cost	Notes
System prompt	~4K tokens	Always, invisible
Auto memory (MEMORY.md)	varies	First 200 lines / 25KB
CLAUDE.md files	varies	User + project + local
Environment info	~300	cwd, platform, git status
MCP tools	small if deferred	Full schemas load on demand via tool search (ENABLE_TOOL_SEARCH)
Skill descriptions	~20-60 each	Bodies load only when invoked

Inspect: /context

/context

Shows a live breakdown of what’s consuming context. Run it when a session feels sluggish or before a big task.

Reclaim Space

Tool	What it does	When to use
/clear	Empty context, fresh start	Switching to an unrelated task
/compact [focus]	Replace history with summary	Mid-task, context near full
/rewind → Summarize	Compress only part of the conversation	Verbose debugging detour you’re done with
Subagents	Verbose work happens in a separate window; only summary returns	Test runs, doc fetching, log analysis

Rule of thumb: /clear between tasks, /compact within a task.

Auto-Compaction & Microcompaction

• When context nears the limit, Claude Code auto-compacts: summarizes the conversation and continues. You’ll see a “context low” warning first.
• Microcompaction silently trims old large tool outputs (file dumps, test logs) before resorting to full compaction.
• What survives compaction: CLAUDE.md, system prompt, memory, the summary. What doesn’t: raw tool outputs, full message history (still in the transcript on disk).

Prompt Caching (why long sessions stay cheap)

Claude Code caches the conversation prefix with the API (5-minute TTL). Each turn only pays full price for new tokens. Implications:

• Rapid iterations are cheap; a session idle >5 min re-reads everything uncached
• Switching models (/model) invalidates the cache — the picker warns you
• Big CLAUDE.md files cost you on every cache miss, not just once

Practical Hygiene Checklist

1. Keep CLAUDE.md lean (< ~150 lines); push detail into skills that load on demand
2. Don’t paste huge files into chat — give paths and let Claude read selectively
3. Ask for targeted reads: “read only the auth middleware section”
4. Delegate log-heavy commands to subagents
5. /context when in doubt; /clear liberally between tasks
6. Watch the status line — configure it to show context usage (see 05-interactive-mode-shortcuts.md)

Real Case: Long Debugging Session Going Stale

Symptoms: slow responses, Claude forgetting earlier decisions.

1. /context                     → 78% used, mostly old test output
2. /rewind → "Summarize up to here" on the point where debugging started
   → early exploration compressed, recent work intact
3. Continue with full precision on the actual fix

Alternative when the task is done: /clear, then start the next task with a one-line recap.

🧠 Quick Recall

• Trick: “Window = RAM, transcript = disk.”
• “Clear between tasks, compact within them.”

👨‍🏫 Teach It

• Show: /context before and after a big file read — make tokens visible.
• They’ll trip on: the 5-minute cache TTL — a long lunch makes the next turn slow (that’s normal).

---

Learning path: ⬅ 08-sessions-and-checkpointing.md · Index · ➡ 10-plan-mode-and-thinking.md