← Back to articlesHarnesses and AI coding

What belongs in a checkpoint for a long AI task?

A checkpoint records the goal, completed and pending work, file locations, constraints and next action.

The short answer

Going well is not enough. Actual state is more useful for resumption than an unstructured chat log.

Try it in three steps

  • Save artifact locations after each stage.
  • Separate passing, failing and untested checks.
  • Provide an immediately actionable next step.

A practice prompt

Before you start

Turn scattered progress into a usable checkpoint.

Practice material (fictional)

draft.txt exists but its contents are unchecked. Input is notes.txt. No sending or publication. Next action: read back and compare.

Replace placeholders in the prompt with this material, or append the material if there are no placeholders. Reading time excludes doing the exercise.

Write a handoff separating historical conclusions from current checks, preserving the no-publication limit and next action without secrets.

What to look for (not a recorded test result)

Preserve unchecked status and filenames so the next session starts with review rather than treating generation as acceptance.

If it goes wrong

If the record has plans but no artifacts, add real paths and status without relabeling plans as done.

Check your result

A new session knows where to resume without guessing.

References and scope

The links below support the concepts or product navigation. The steps and exercises are our teaching suggestions, not official certification or measured customer outcomes. Product interfaces and availability may change. References checked on 2026-10-17.

Start with one task

Describe the current process and its slowest step. The initial conversation is free.

Let’s talk ↗
Need help? See small-scale implementation services ↗

Keep reading