Claude Code: keeping the context window healthy

The Claude Code context window is finite. Every message, every tool call, every file read takes a slice of it. When it fills up, two things happen: the model starts forgetting what was said early in the session, and per-turn costs jump because processing 200k tokens each round isn’t free.

Four habits I use daily to keep it healthy.

/clear, /compact and auto-compact

Three ways to free up context, each for a different situation.

/clear wipes everything and starts a fresh session. I use it when switching tasks: the payments bug is done, now I’m touching logs, I don’t need the conversation dragging the previous thread along.

/compact asks the model to summarize the context and keeps only that summary. Useful after an hour of work, when the 12 files I read at the start are no longer pulling their weight but I do want to keep the decisions we made along the way.

Auto-compact fires on its own when you hit ~95% of the window. It works, but it can catch you mid-task: the automatic summary may drop nuances you were about to lean on. If I see the indicator climbing past 70% and the task isn’t closed, I trigger /compact myself before the model decides for me.

Subagents to keep the main session clean

When a task starts with “I need to understand how X talks to Y”, that exploration phase easily burns 30-50k tokens: greps, finds, reading several files to reconstruct the flow.

Instead of doing that in my main session, I launch an Explore subagent. The subagent has its own window: it spends its tokens exploring and returns a half-kilobyte summary. My session only receives the summary, not the dozen full files.

I start the task with the main context almost empty, already holding the mental map I needed, and the window stays free for the real work — writing, reviewing, iterating.

@path: hand over the file, don’t ask Claude to find it

Two ways to get a file into context: ask Claude “read the tests to figure out what X does” (Claude lists, opens, sometimes picks the wrong file), or type @path/to/test.go yourself.

The second is direct. The file is injected and that’s it. No searches, no intermediate tool calls, no back-and-forth.

Same idea with the ! shortcut: !cat config.yaml drops the command output into context without going through the model at all.

Claude searching and reading ten files to find the right one costs more than just handing over the two that matter from the start.

Plan mode and when to start a new session

Plan mode (Shift+Tab twice) makes Claude design without executing. No tool calls, no long outputs, no 800-line logs dumped into the conversation. Just reasoning and a plan at the end. The design phase stays cheap, and once you execute you already know what to touch.

To decide between continuing and starting fresh:

• New task, same project → /clear. Keep CLAUDE.md and the local setup, drop the conversation.
• Want to explore an alternative without losing the current state → fork the session. Two branches in parallel.
• Completely different task → new session from scratch.

Starting clean has way less friction than fighting a bloated session.

Why this matters

The window is a budget, not an infinite. Every read, every tool call, every subagent has a cost, and answer quality drops as you approach the ceiling.

If you notice late, /compact bails you out. If you notice in time, you don’t even need it.