Nano Banana Pro 缩略图的高级两步提示策略

这条推文概述了一种先进的两步提示策略，可显着提高人工智能输出的质量，特别是使用 Nano Banana Pro 生成 YouTube 缩略图。该策略首先要求强大的法学硕士（如 GPT-5.2）系统地总结创建 Nano Banana Pro 视觉概念（创建“地图”）的最佳实践、术语和约束。然后，用户要求法学硕士根据已建立的系统和用户的要求生成最终的、具体的视觉概念，从而产生更加一致和有针对性的输出。

英文原文

Advanced Two-Step Prompting Strategy for Nano Banana Pro Thumbnails

This tweet outlines an advanced, two-step prompting strategy to significantly improve the quality of AI output, specifically for generating YouTube thumbnails using Nano Banana Pro. The strategy involves first asking a powerful LLM (like GPT-5.2) to systematically summarize the best practices, terminology, and constraints for creating Nano Banana Pro prompts (creating a 'map'). Then, the user asks the LLM to generate the final, specific prompt based on that established system and the user's requirements, resulting in much more consistent and targeted output.

第一次写图片提示词？不用理解英文变量，按照下面的中文步骤补充画面要求即可。

适合用途: YouTube 封面
建议模型: Nano Banana Pro
参考素材: 无需上传

开始填写这个模板

Actually, there's a trickier way to use this. First, have it systematically summarize the information → Then, based on that, have it create the output/deepen the discussion Just doing this makes the quality of the output completely different. This is because if you first create a blueprint in the AI's mind—including "terminology," "judgment criteria," "prohibited items," "patterns," and "priorities"— The next instructions will be output according to that blueprint, making it less likely to deviate. For example, if you want Nano Banana Pro to create a thumbnail: ① First, ask GPT-5.2: "Thoroughly research and systematize the best practices for creating prompts for YouTube thumbnails using Nano Banana Pro. Include common mistakes as well." ② After that, ask: "Based on the current system, create a prompt for making a YouTube thumbnail on the theme of XX using Nano Banana Pro. The video content is like this, and I want to include this kind of wording." Then, input the completed prompt into Nano Banana Pro. It becomes much easier to get the desired output. If you just ask it to "create" every time, the AI outputs based on the moment, leading to inconsistent results. But if you create a map (system) first, it's powerful. Moreover, this method works for side business strategies, sales copy, and video scripts. The versatility is insane. I highly recommend having the AI organize how to create good output before having it perform the task, and then having it think based on that organization. By the way, the thumbnail created this time is the third image. Since the first attempt was mediocre, I had it systematize how to create attractive YouTube thumbnails before generating the prompt.

Actually, there's a trickier way to use this. First, have it systematically summarize the information → Then, based on that, have it create the output/deepen the discussion Just doing this makes the quality of the output completely different. This is because if you first create a blueprint in the AI's mind—including "terminology," "judgment criteria," "prohibited items," "patterns," and "priorities"— The next instructions will be output according to that blueprint, making it less likely to deviate. For example, if you want Nano Banana Pro to create a thumbnail: ① First, ask GPT-5.2: "Thoroughly research and systematize the best practices for creating prompts for YouTube thumbnails using Nano Banana Pro. Include common mistakes as well." ② After that, ask: "Based on the current system, create a prompt for making a YouTube thumbnail on the theme of {argument name="theme" default="XX"} using Nano Banana Pro. The video content is like this, and I want to include {argument name="wording" default="this kind of wording"}." Then, input the completed prompt into Nano Banana Pro. It becomes much easier to get the desired output. If you just ask it to "create" every time, the AI outputs based on the moment, leading to inconsistent results. But if you create a map (system) first, it's powerful. Moreover, this method works for side business strategies, sales copy, and video scripts. The versatility is insane. I highly recommend having the AI organize how to create good output before having it perform the task, and then having it think based on that organization. By the way, the thumbnail created this time is the third image. Since the first attempt was mediocre, I had it systematize how to create attractive YouTube thumbnails before generating the prompt.