Practical prompt patterns for vision-language models: constrained outputs, role-scoped system prompts, chain-of-thought reasoning, and a VLM benchmark comparison framework.