What you’ll know by the end of this check

What’s actually in your context window — and what it costs
When to /compact, when to /clear, and when neither is right
Three habits that stretch your context 2–3x without losing fidelity

The shortest possible answer

Context is Claude’s working memory. Every file it reads, every command output, every message you type — all of it takes space. Manage it deliberately or quality silently degrades.

Three commands you should know cold:

Command	What it does	When to use
`/context`	Shows usage breakdown — what’s taking up space	Always. Run it when you notice Claude getting sloppy.
`/compact`	Summarizes the session, keeps the memory	You’re deep in one feature and need to keep going
`/clear`	Wipes the context completely	Switching to a new feature or a new problem domain

The auto-compaction failure mode

When you approach the context limit, Claude Code auto-compacts. It summarizes older content and removes tool-call noise. This is not free — subtle details can get lost.

If you catch yourself thinking “Claude used to know this,” check whether auto-compaction ate it. Put the fact back in your CLAUDE.md so it survives the next compaction.

Three habits that buy you context

1. Be specific

A vague prompt looks smaller but costs more. Without concrete instructions, Claude has to explore your codebase and reason its way to what you meant — and that exploration is what actually fills the window. One paragraph of specifics beats one sentence of vagueness.

2. Manage your MCP servers

MCP servers load all their tool definitions into context the moment they connect — even when you’re not using them. If you have five MCP servers configured but only need one, the other four are paying rent in your window.

Run /mcp to see what’s connected. Disable anything you’re not using this session.

3. Use subagents

Subagents run in parallel with their own context window. For “just give me the answer” tasks — “find the authentication code in this repo”, “summarize the failing test output” — a subagent does all the exploration in its own window and hands you back only the summary.

Your main context stays clean.

When NOT to compact

If you’re about to make an irreversible decision (ship, deploy, approve a large refactor), do NOT compact mid-decision. Compaction can drop the exact detail that matters. Finish the decision first, then compact.

Things to try right now (10 minutes)

In an active session, run /context. Screenshot the output.
Identify the two biggest consumers of your window. Write them down.
Do one of: disable an unused MCP server, move a recurring fact into CLAUDE.md, or close a subagent you spawned but never read.
Run /context again. See the reclaim.

The canonical version

Full official lesson is at anthropic.skilljar.com/claude-code-101/469793.

Ready to verify this check?

You can read /context output and name the top consumers. You know the difference between /compact and /clear and when to use each. You’ve reclaimed context at least once. Mark it cleared.