They should have also added some basic chain-of-thought reasoning in the prompt + asked it to add [nitpick] tag at the end, so that the likelihood of adding it

They should have also added some basic chain-of-thought reasoning in the prompt + asked it to add [nitpick] tag at the end, so that the likelihood of adding it correctly increased due to in-context learning. Then another pass removes all the nitpicks and internal reasoning steps.