A compliance platform runs thousands of regulatory document reviews per week. The team observes that Claude's performance degraded significantly after migrating from explicit numbered instructions to a prose-style system prompt. The original prompt averaged 95% accuracy on extraction tasks; the prose version drops to 78%. Both prompts contain the same information. The team wants to restore accuracy without reverting to numbered lists exclusively. Which prompt engineering principle most directly explains the degradation, and what structural fix addresses it?
Select an answer to reveal the explanation.
Short Explanation
Here's the deal — prose buries your extraction constraints, so the model misses individual rules. Numbered lists work because they create explicit attention anchors. Use XML tags to wrap each rule as a discrete unit without going back to numbered lists.
Full Explanation
Prose burial of constraints is the most direct explanation for extraction accuracy degradation. In structured extraction tasks, the model must reliably apply each specific rule to each specific input element. Numbered lists create discrete, identifiable rule units and explicit attention anchors that the model can check against systematically during generation. When the same rules are embedded in flowing prose, they merge with surrounding text; the model may process the prose holistically and miss applying individual rules precisely, reducing the salience of specific extraction rules. XML tagging, such as <rule id='1'>...</rule> or <extraction_rule id='account_number'>...</extraction_rule>, restores the discreteness of each rule without forcing numbered list formatting, allowing prose narrative context while maintaining rule salience. The token-count/context-ratio explanation is plausible but the token count difference between numbered lists and equivalent prose is typically minimal, insufficient to explain a 17-point accuracy drop. The explanation that prose system prompts are treated as background context rather than imperative instructions, with the fix being to move all extraction rules into few-shot examples, is not the most direct explanation or fix. Moving all extraction rules into few-shot examples changes the nature of the instruction from declarative to demonstrative, which is a valid alternative approach but a much larger change than needed, and few-shot examples for complex regulatory extraction require many examples to cover all rule types. The priority-conflict explanation is plausible for ambiguous inputs, but extraction accuracy drops across all inputs, including unambiguous cases, indicating a general rule-application problem rather than only conflict resolution.