300字以内 vs write all: order and conflict rule shift LLM output

The test grew out of a common editing habit: writing a prompt, then remembering something else and appending it at the end. The team had a prior belief that such trailing additions carry less weight, and they also drew on the standard prompt-engineering advice to specify priority when constraints conflict. Their web search turned up reports of a primacy bias — roughly 73% of cases favored information at the start of a prompt — and a paper on incompatible instructions (PRIME: Evaluating Prompt Resolution Under Incompatible Instructions in LLMs, arXiv:2606.22470). Based on that, they expected Condition A, where the character limit came first, to be obeyed most strictly.

Condition A did stay well under the 300-character ceiling, but finished at 135 characters rather than using the full budget, which the team found notable. Condition B, with the completeness demand first, came in at 237 characters. The surprise was Condition C: identical word order to A, yet it produced 233 characters, nearly matching B. If simple position bias explained everything, C should have looked like A.

The team's reading is that the difference lies in the wording of the appended instruction. Condition A only demanded that no reason be omitted; Condition C added a conflict rule saying what to cut. In A, the model may have treated the two demands as simply incompatible and defaulted to the first constraint; in C, it may have read the rule as permission to attempt both. The practical implication they draw is to state both constraints together from the start and include the resolution step — though they caution this is a post-hoc explanation from one run per condition.