Add Checkpoints to Long Tasks: Make AI Submit Intermediate Saves at Every Step

When you hand a long task over to an AI, the most common failure mode looks like this: You wait twenty minutes for the final deliverable, only to open it and realize it went off the rails in step two. Sending it back for revisions or starting from scratch are both costly losses.

The checkpoint technique is the cure for this. Instead of demanding that the AI complete everything in one go, require it to pause at the end of each stage and submit a structured intermediate output: what was done, what decisions were made, what remains, and where your input is needed. You are reviewing the *output*, not the work process.

Specific Scenarios

**Scenario 1: Long-form Content Writing.** For a 3,000-word video script, don’t just tell it to "write the whole thing." Break it down into checkpoints: First, produce an outline (including core arguments for each section) → You approve → Then, draft key paragraphs → You approve → Finally, produce the full text. If the outline is wrong, you save the effort of fixing the entire text; if the direction is right, polishing the full text costs very little.

**Scenario 2: Code Migration.** When migrating a project from Python to TypeScript, checkpoints could include: a mapping table of function signatures → a list of completed files with descriptions → self-test results. By reviewing just the signature table and test outputs, you can catch 80% of the issues.

**Scenario 3: Weekly Reports or External Materials.** First, have the AI output a list of facts, citing the source for every number. Once you confirm the facts are correct, let it start writing. This is much safer than letting it improvise freely and then trying to spot errors.

Copy-Paste Prompts

"Break this task into three nodes. After completing each node, stop and list: 1) Completed steps (with evidence); 2) Decisions I made and why; 3) Potential error points; 4) Whether I need to make a decision. Do not proceed to the next step until I reply."

Note that the keyword is **"evidence."** "Completed" must include file paths, command outputs, or data sources. A simple "It's done" is not acceptable.

When to Use It, and When Not To

**When it’s worth using:**

- The task has more than three stages, or the total estimated time exceeds ten minutes.

- The cost of modifying the final deliverable is high: public-facing content, client documents, or migration scripts running in production.

- You are unsure about the AI’s direction: new domains or unfamiliar tech stacks.

**When not to use it:**

- Short tasks that can be completed in one shot: naming something, fixing a typo, or fact-checking.

- Stages where you cannot verify the result: Asking for approval on color schemes for design mockups is meaningless if you can’t judge them. The prerequisite for checkpoints is your ability to verify.

- Exploratory phases with no defined path: Brainstorming should run free. Checkpoints are for "verifying the process," not for "exploring directions."

Pre-flight Checklist

- [ ] The task can be broken down into 2 to 4 checkpoints, and you know what deliverable is expected at each.

- [ ] You are willing to spend two or three minutes verifying each checkpoint. If verification is more expensive than rework, this task isn’t suitable for checkpoints.

- [ ] You have defined the output format: Evidence, Decisions, and Risks—all three must be present.

Common Pitfalls

1. **Checkpoints turn into criticism sessions.** If you only reply "OK" every time, checkpoints merely lengthen the conversation and bloat the context. Write down your feedback during verification and have it fixed on the spot; don’t save everything for the end.

2. **Outputs become increasingly messy.** It’s recommended to save the output of each checkpoint as a separate file, then reconcile them item-by-item against the final deliverable. This record can also serve directly as handover documentation.

3. **Verification relies on gut feeling.** Glancing only at the titles is as good as not verifying at all. Force yourself to spot-check at least one piece of evidence: open that file, run that command, or verify that data source.

4. **Checkpoints are set too granularly.** The purpose of checkpoints is to "catch directional errors early," not to "micromanage every detail." Granularity should be based on stages. If a stage is wrong, you can still change direction; obsessing over details just exhausts you.

The core principle in one sentence: Don’t manage the AI’s labor process; manage its decision nodes. Require evidence at each node, and you verify that evidence.