The checklist test: how children learn to verify instead of trust
A checklist bolted onto a skill is the difference between hoping an AI's output is right and knowing it. Here is how a simple pass/fail check teaches a child to verify instead of trust.
A student's project-writeup skill looked finished - six tidy sections, correct headings, confident language. We added one check to it before letting him submit: "the method section must contain at least one actual measurement." It failed immediately. The section that looked most complete had no numbers in it at all, just well-written sentences about what he had done.
Nobody had lied to him. The output simply looked finished without being finished, and looking finished is not the same thing as being correct. That gap is exactly what a checklist exists to close.
How do I check AI output is correct?
You attach specific, checkable rules to the skill itself - not vague hopes like "make sure it's good," but pass-or-fail conditions a child can actually test against. Does it have exactly three sections? Does the answer include a number? Does every claim have a source line next to it? A rule that cannot be answered yes or no is not a check.
- "Must have exactly three characters" - checkable in ten seconds by counting.
- "Must include one measurement in the method section" - checkable by scanning for a number.
- "Must not use any word longer than the reading level allows" - checkable, if slower.
- "Must be good" - not a check. It is a hope wearing a check's clothes.
Why this is the difference between trusting and verifying
Trusting an AI's output means reading it, feeling like it's probably fine, and moving on. Verifying means running an actual test against it and getting a real pass or fail. Those feel similar in the moment - both end with the child accepting the output - but only one of them would have caught the missing measurement.
Checks are how you stop trusting and start verifying. The feeling of confidence is not evidence.
Building the habit early
What a checklist cannot catch
A checklist only tests what it was written to test. "Has a measurement" does not confirm the measurement is correct - only that one exists. A thorough checklist reduces the ways an answer can quietly be wrong; it does not eliminate them. That honest gap is why the next level, connectors, spends its first article on permissions rather than capability - because a skill that checks its own output still has to be trusted with what it is allowed to touch, which is a separate question entirely.
Questions we get asked
How do I check AI output is correct?
Attach specific, pass-or-fail rules to whatever the AI produces - a required number of sections, a mandatory measurement, an exact word count limit - rather than judging by whether it feels finished. A rule that can be answered with a clear yes or no is a real check; a vague hope that it "seems right" is not.
What is the difference between trusting and verifying AI output?
Trusting means reading an answer, feeling reassured by how confident and complete it looks, and accepting it. Verifying means running an actual, specific test against the output and getting a real pass or fail result. Confident-sounding output can still fail a real check, which is why the feeling of correctness is not the same as correctness.
Can a checklist catch every mistake in AI-generated work?
No - a checklist only catches what it was specifically written to test for. A check confirming a measurement exists does not confirm the measurement is accurate. Checklists reduce the ways an answer can be quietly wrong; they do not remove the need for judgment on anything the checklist did not think to test.