Abstract
This paper examines structural prompt engineering for Claude, a large language model trained by Anthropic with Constitutional AI (CAI). We formalize a four-component procedure, Constitutional AI-Aligned Structured Prompting (CASP), comprising explicit role and context framing, XML-based input structuring, explicit step-by-step reasoning, and output control through schema-constrained structured outputs, with prompt caching for reused contexts. CASP is specified as an algorithm and evaluated in a controlled ablation study on two public benchmarks: entity extraction to JSON on CoNLL-2003 (T1; 50 items, also run with an embedded conflicting instruction) and multi-step arithmetic on GSM8K (T2; 50 items), using Claude Sonnet 4.6. The full procedure, a naive baseline and leave-one-out ablations were compared using automatic gold-based scoring and Holm-corrected paired Wilcoxon tests. Full CASP achieved a typed entity F1 of 0.855 on T1 versus 0.000 for a naive baseline that did not specify an output format (p < 0.001), because none of the baseline's prose answers was machine-parseable; when these prose answers were parsed post hoc, CASP remained 0.097 higher (p = 0.005). On T2, all conditions scored 0.960 except when explicit reasoning was removed (accuracy 0.740, p = 0.017, r = 0.846), because a response schema without a reasoning field suppressed the model's spontaneous reasoning. XML showed no significant advantage over Markdown. No condition actually obeyed the embedded instruction. Prompt caching reduced input cost by 85% with no significant change in output quality (p = 1.000, output agreement 0.93). Code, sampled data and raw outputs are released for replication.
