/clear vs /compact — pick the right reset

Long Claude Code sessions get expensive linearly. Every turn re-sends the entire conversation as input. By turn 40, you might be paying for 40,000+ input tokens every keystroke just to keep history around.

/clear and /compact are the two ways out. They do different things.

/clear — start fresh

Wipes the conversation history entirely. Claude has no memory of the prior turns; only the persistent CLAUDE.md + any new files it reads. Use it when:

• You've finished a discrete task and the next one is unrelated.
• The conversation drifted off the original goal and the history is now noise.
• You're paying >$0.10 per turn just to keep talking.

The downside: any context Claude had on what you were trying to do is gone.

/compact — summarize and continue

Hands the conversation history to Claude with one instruction: write a tight summary of what's relevant for continuing. Then replaces the history with that summary. You keep continuity at a fraction of the cost. Use it when:

• You're in the middle of a multi-step task and want to keep going.
• You've made decisions in this session that the next turns need to know about.
• The conversation has bloated past 30k tokens but you're not done.

The downside: the summary is lossy. Claude might lose a constraint that wasn't obviously important to preserve. Re-state critical invariants in your next prompt.

When to fire which

Situation	Use
Task finished, new task	/clear
Same task, getting expensive	/compact
Same task, just got confused	/clear + re-prime
Working on three things in parallel	/clear and run them in separate sessions

The cost trigger

A reasonable rule of thumb: if your last 5 turns each cost more than $0.05 and your task isn't visibly progressing, it's /clear time. The Session Cost Tracker (live via the CLI hook) shows this number in real time.

The /clear-/compact Coach module turns this into a per-session recommendation.