A city needs Claude to cite ordinance sections instead of vague paraphrases. Which prompt engineering step fits?

Writing clearer system instructions so Claude cites ordinance sections instead of paraphrasing vaguely is the correct prompt-engineering step because citation behavior is instruction-sensitive. Cities need answers that point residents and clerks to enforceable section numbers, not impressionistic restatements that cannot be verified at the counter. Explicit system instructions—require section IDs, quote short operative clauses, and refuse when the corpus lacks a cite—steer Claude toward grounded municipal guidance.

Appending full ordinance excerpts to the user prompt while leaving the existing system guidance unchanged fails because it dumps raw text into the conversation without fixing the underlying system instructions, so Claude still won't reliably cite section numbers and the prompt just gets bloated. Raising the sampling temperature to encourage more detailed ordinance section references fails because higher temperature adds randomness and makes citation style less predictable, increasing the risk of hallucinated or inconsistent section numbers rather than grounding them. Adding step-by-step reasoning requests to the user message fails because it asks Claude to reason through citations in the moment without changing the system-level grounding rules, so it still won't reliably require section IDs or refuse uncited answers.

Exam caveat: clearer instructions cannot invent sections that are missing from the retrieved corpus; pair instruction clarity with retrieval of the actual ordinance text or the model may fabricate plausible-looking cites. Operational check: A/B the vague vs citation-strict system prompts on a fixed ordinance FAQ set, score presence and correctness of section numbers against the official code, and promote the prompt only when citation precision clears the city’s published threshold.