Battle-tested system prompts & few-shot schema directives guiding LLMs (GPT-4o, Claude 3.5, Gemini) to return strict, machine-parsable JSON.