Sources checked
Give the evaluator concrete criteria
An original article check can ask whether the opening answers the headline, whether claims have sources, and whether a comparison preserves important limitations. Those checks are more useful than asking a second model whether it likes the draft.
Feedback models can approve the same mistake
A model can approve another model’s mistake. Use programmatic checks for facts that can be computed, and human review where the evaluation depends on judgment or source interpretation.
What the critic is allowed to know
The pattern needs two roles. The drafter produces a candidate.
The evaluator says whether the candidate meets a stated test, and why. If both roles are the same prompt, the second pass often repeats the first opinion in new words.
Give the evaluator the rubric and the candidate, and withhold the drafter’s excuses.
The rubric has to be checkable. “Be more insightful” produces longer text.
“Every number must appear in the source paragraph” produces a yes or no. Prefer checks that code can do: schema, citation present, total matches the rows.
Use a model as the evaluator only for judgments you have sampled against a person.
When to stop
A loop without a budget will spend until the evaluator is bored into a pass. Cap the rounds.
If the evaluator still fails the last draft, return the failure to a person or to a fallback. Do not ship the last draft just because you ran out of patience and the prose looks finished.
Measure whether round two actually fixes round one on a set you labeled. Some tasks get worse: the drafter “fixes” a correct sentence to satisfy a vague critic.
If the labeled set does not improve by the second round, the loop is a cost, not a pattern to keep.
What to record before you rely on it
Write the rubric as items a stranger could score. Record the pass rate after one draft and after the capped loop, on the same inputs.
Save a case where the evaluator was wrong, so the rubric gets edited instead of the loop simply running longer.
Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test. When the result is a failure, store the artifact and change the design.
Do not relax the check until the model happens to pass. Name a person who owns that failure.
A note with no owner and no inputs cannot be repeated, and a pass with no date cannot be compared with the next model id.
Common questions
Can the same model be both drafter and critic?
It can, if the critic prompt sees a rubric and the draft, not a request to defend the draft. Check the critic against people.
Being a second call does not make it independent in judgment.
Is this the same as a longer prompt?
No. A longer prompt still produces one draft. The loop spends extra calls in exchange for a chance to repair a specific failed check.
If you cannot name the check, lengthen nothing.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





