Ask for evidence at the next checkpoint

A long answer can sound complete while leaving the important question unanswered: what changed, and how do we know? At a checkpoint, ask for a small evidence record before deciding whether another round of work is useful.

This is an original workflow template. It has not been benchmarked against other prompts, and it does not establish that a particular model is more reliable.

Start with the intended result

Write one sentence describing the result you actually need. For a fictional website task: “A visitor can submit the contact form and the owner receives the enquiry.” A button animation or a successful build supports only part of that result.

Keep the acceptance condition visible when you ask for progress. Otherwise, a detailed account of implementation can gradually replace the original goal. The checkpoint should reconnect the current artifact to that condition.

Ask for observations, not confidence

Try this prompt:

Compare the current result with the acceptance condition. List what you directly observed, what remains an assumption, and the smallest check that would resolve the most important uncertainty. Include the artifact or record behind each observation.

“Directly observed” still needs scrutiny. A model can misread a log or describe a check it never ran. Open the referenced artifact and inspect whether it supports the specific claim. For a form, a local test response is different evidence from receipt in the intended destination.

Keep a compact checkpoint card

Use five fields: intended result, current artifact, evidence, unresolved question and next action. Add an owner if another person must make the decision. Keep private customer information out of a shared example.

In the fictional form task, the card might say that validation works locally but delivery has not been checked. The next action is then a controlled delivery check, not another visual redesign. That makes the limitation actionable without pretending the whole task failed.

Choose one intervention

The Keep Working deck can supply a short nudge when you already know the kind of intervention needed. Its core categories include making a result concrete, stepping back, critique and finishing. The category API exposes the current catalog; the app serves text cards and does not run your project or verify the answer.

Choose a card that addresses the uncertainty on the checkpoint. Repeatedly asking for more confidence is less useful than specifying the missing observation. If the evidence is already sufficient, the next action can be to finish rather than to keep prompting.

Define the stopping point

Close the checkpoint when the agreed condition is supported, when a named blocker needs a decision, or when the remaining work no longer serves the goal. Record which case applies.

For a reusable handover, save the checkpoint card with the artifact. For testing the wording itself, use the small prompt comparison worksheet. Neither exercise turns a generated explanation into proof; it gives the reviewer a clearer place to look.