checkpoint

Save and restore working state across AI context boundaries.

When a long session approaches the context window limit, the working state — what you were doing, what files you'd read, what the next step was — gets compressed or lost. The next session starts cold and has to reconstruct everything.

checkpoint gives Claude a place to save that state before the cliff hits, and load it back in the next session.

Install

uv tool install git+https://github.com/benthomasson/checkpoint
checkpoint install-skill

How It Works

Claude writes .claude/checkpoint.md directly when saving — the file contains task, status, key files, next step, and context. There's no checkpoint save command; the skill teaches Claude to write the file itself.

checkpoint load        # display current checkpoint (with age)
checkpoint clear       # delete after task is complete
checkpoint init        # create a blank template
checkpoint install-skill   # install Claude Code skill

The Skill

After running install-skill, Claude will:

• Save proactively when context is getting long — without being asked
• Load at session start when a checkpoint exists, orienting immediately
• Write specific, actionable checkpoints ("next step: update SKILL.md then commit") not vague ones ("working on beliefs tool")

Part of a Toolkit

Pairs with:

• entry — time-ordered documentation entries
• beliefs — belief registry with staleness detection