Skills
Give agents structured markup to read and plain prose to write, because XML instructions cut answer quality from 74% to 48%
A deployed tender-response system found the formatting rule is asymmetric: structured markup helps where the model reads source documents, but converting the instruction material itself from prose to nested XML dropped answer quality from 74% to 48%. Worse for anyone maintaining a style guide, explicitly naming forbidden constructions increased how often the model produced them. An LLM judge rated the system's answers at least as good as the human-submitted answer on 40 of 55 ground-truth sections, and of 15 adverse verdicts only 6 were writing-quality failures the system could have avoided.
↳ Follow the thread