Do XML tags work better than JSON formats for long prompts?

Going from a 15% error rate down to under 2% just by re-ordering instructions is impressive. Context drift is such a headache when handling variable article lengths.

Putting data first and instructions last definitely forces the attention mechanism to prioritize the output specs. I usually rely on markdown headers and line breaks to segment my prompts, but structuring with custom tags seems much cleaner.

Do you apply this same pattern when chaining multiple prompts together?