开启时会模糊预览图，关闭后正常显示
RT @thegarrettscott: .@pipedream_labs is a proud customer of RMFG. No one beats their customer service and desire to help you win.
making complex metal assemblies for fast moving companies | Founder at RMFG
Got a chance to spend a bit of time hearing @67Designs American manufacturing story this morning. One of the people that care most deeply about this problem Thanks @67Designs for the time
making complex metal assemblies for fast moving companies | Founder at RMFG
i'll be at my favorite crypto conference, @solana breakpoint, next week if you're building a new app or anything that accelerates the adoption of Internet Capital Markets, let's hang
investing @a16zcrypto | winning is everything
1. https://t.co/M553AhfmSu 2. https://t.co/E7FoNQQ9zE 3. https://t.co/rskFKI6BEZ 4. https://t.co/hL0hzVaTmd
💻 https://t.co/Y30jsaHwz9 $30K/m ⚡️ https://t.co/vatLDmi9UG $21K/m 📈 https://t.co/3EDxln5mdi $17K/m ⭐️ https://t.co/MZc8tG9xWi $17K/m 🍜 https://t.co/r07EpGSYJ2 $1K/m 🧬 https://t.co/SfrVXVtmdA $0/m 🧾 https://t.co/7olaOzV8Xd $0/m +20 https://t.co/4zCWHGJp1S
Here are the principles and methods I use when writing prompts: My Prompting Principles 1. Design templates, not fixed prompts. I want people to adapt the structure to their own scenarios—whether it’s their favorite brand, city weather, stock market themes, mini pop-up stores, or turning an article into an infographic. The key idea is to give users a flexible framework rather than locking them into a single use case. 2. Leverage the model’s native capabilities (search, world knowledge, reasoning). For example, the city-weather prompt lets Gemini fetch real weather information for a given city and date. The 3D mini stock-market prompt works similarly. Infographic prompts rely on the model’s comprehension and summarization of the input article. My Strategy for Building Nano Banana Pro 1. Start by using AI to build a working prototype for one specific scenario. 2. Then generalize it into a prompt template that can dynamically adapt based on user input. The core idea is to let the model assemble the structure automatically and adapt to different scenarios without imposing artificial constraints, just like writing code, where you avoid hard-coding and instead keep flexible interfaces for composition. About Prompt Length Modern models are strong enough that token length isn’t a big concern. I don’t spend much time optimizing or shortening prompts at the beginning—functionality comes first. If needed, refinement can come later. I mainly use GPT-5.1, GPT-4.5, and Gemini 3 Pro as my supporting tools. I send the same task to all of them in parallel, let each model generate its own output, and then choose whichever result turns out best.
prompt for the cover image
我在写 prompt 时的原则和方法大概是这样： 我的提示词原则 1. 做成模板，而不是固定提示词。 这样每个人都可以根据自己的场景自由发挥——输入自己喜欢的品牌、城市天气、公司股市、迷你店主题，甚至把文章做成信息图。核心是给用户一个“能玩”的框架，而不是规定他们必须怎么写。 2. 充分结合模型自身的能力（搜索、世界知识、理解能力）。 比如城市天气 prompt 可以让 Gemini 根据城市与日期检索天气；迷你 3D 公司股市图也是利用模型的实时金融检索；信息图 prompt 则依赖模型对文章内容的理解与提炼。 我写 Nano Banana Pro 的策略 1. 先用 AI 针对一种具体情况跑通原型。 2. 再把这个原型抽象成一个可扩展的提示词模板，让输入内容决定最终输出。 核心是让模型自动组合结构、自动适配场景，不去人为限制它。就像写程序，不 hardcode，而是留好接口灵活组合。 关于精简提示词 现在的模型能力很强，token 长一点或短一点影响不大。 所以我不会过度花时间去压缩 prompt，而是优先确保它“能用、能生长”。 等功能成熟后，必要时再考虑精简。 我主要使用的辅助工具是 GPT-5.1、GPT-4.5 和 Gemini 3 Pro，会把同一个任务同时发送给它们，让它们各自生成结果，再从中挑选表现最好的。
本文封面图提示词