Three-layer approach keeps AI-generated questionnaires structurally valid

On November 6, 2026, a developer posted an article on Juejin, a Chinese developer community, titled "Making AI Reliably Output Parsable Questionnaires: schema Constraints, Retries, and Degradation," describing a three-layer approach for keeping AI-generated questionnaire output parseable. The starting point was a batch-generation task for course evaluation questionnaires: the author fed two hundred requirement descriptions through the model in a single pass, and a spot check before collection turned up thirty malformed results. Some had option lists misaligned with their question stems, some had skip rules pointing to question numbers that did not exist, and some came back as half-finished JSON. Rather than silently shipping these, the author proposed a pipeline built from three layers. First, a schema constrains the model's output to a fixed, parseable structure, so any result that deviates from the expected shape can be rejected before it reaches the user. Second, generation failures trigger automatic retries, giving the model another chance to produce valid JSON under the same schema. Third, when retries still do not yield a usable result, a fallback degrades the output gracefully instead of letting a broken questionnaire through. This is presented as practical engineering experience rather than a general benchmark or product launch, and the article does not cover broader claims about AI question-generation accuracy. The piece is aimed at developers who need to turn natural language requirements into structured questionnaires programmatically, with the three-layer design meant to close the gap between what the model attempts and what downstream systems can actually consume.