一个提示词，让 Opus 5.5 可视化《看不见的城市》
做法与步骤
- 作者长期做数据可视化和可探索式解释，LLM 对他既是福音也是诅咒；GPT-6 Astra 在谜题上是一次跃升，而 Opus 5.5 的跃升体现在设计上。
- 他给 GPT-6 Astra 和 Claude Opus 5.5 同一个提示词：用 three.js（pnpm）可视化卡尔维诺《看不见的城市》全部 55 座城市，一次性任务、不许提问、有 6 小时做到杰作。
- GPT-6 Astra 在 Codex 中端到端跑通，但产出带有 AI 设计垃圾：堆砌概念和注释、未检查是否必要、视觉噪音多，还有本用于无障碍而非直接展示的废话。
- Opus 5.5 在 Claude Code 中只用了 1 小时 25 分钟（作者本要求 6 小时），它自称用了约一半时间、剩余打磨收益递减；作者认为可算作代理时间膨胀：6 个子代理并行约 7 代理小时。
- 作者对 Opus 5.5 的结果着迷，认为即便这已是天花板也很高，并确信下一代模型会更高；他鼓励大家用同一提示词测试不同模型或不同 harness。
- 作者预期 AI 将成为设计中被广泛使用、消耗大量 token 预算的工具，同时自问：在创作交互媒体时，我的位置在哪里？
对实践的启示
对做 AI 应用和用 AI 干活的人：一次性长时提示词已能产出高质量交互设计，值得用同一提示词横向对比不同模型和 harness；同时要警惕 AI 设计垃圾（堆砌概念、视觉噪音），并思考自身在创作流程中的定位。
作者承认的反方与局限
作者承认 Opus 5.5 的结果可能仍被批评为“过于丰富”、缺乏极简感；也承认它没有按要求用满 6 小时，本应把时间用尽。
原文摘引
GPT-6 Astra was a jump when it comes to puzzles. Claude Opus 5.5 is a jump when it comes to design.
Make a three.js (pnpm) visualization of all Invisible Cities by Italo Calvino. Don’t ask questions, it is a one-shot task. You have 6h of work, use it until it becomes a masterpiece.
I’m in awe, but I’m also asking myself: what is my place in creating interactive media?
▸ 展开原文全文（26 段，抓取自 Hacker News 首页）
GPT-6 Astra was a jump when it comes to puzzles. Claude Opus 5.5 is a jump when it comes to design.
Vibe designing
Even local models can make a competent landing page for a shop. Doing interactive visualization is a different tier. Since I am into creating data visualizations and explorable explanations, LLMs were both a blessing and a curse.
Then for a moment I was happy-ish with GPT-6 Astra. My subjective experience was that it is a bit better than Fable 5.1 at a general overview, following intentions behind prompts, and checking that it all works correctly.
Then there was the Opus 5.5 moment for AI-assisted design. I saw an optical explorable explanation:
One may argue that it is still “too rich”, and has no sense of minimalism. But still, wow! I was still in disbelief. Was this really its consistent quality for a one-shot experiment?
Instead of my beloved optics, I went for something different.
Invisible Cities
What’s a good prompt? Well, I went with the beautiful urbanistic poetry of Invisible Cities by Italo Calvino, presenting 55 imaginative cities, each one an emotion or state of mind, expressed in its architecture and in how people behave.
When a man rides a long time through wild regions he feels the desire for a city. Finally he comes to Isidora, a city where the buildings have spiral staircases encrusted with spiral seashells, where perfect telescopes and violins are made, where the foreigner hesitating between two women always encounters a third, where cockfights degenerate into bloody brawls among the bettors.
In 2019, I had a small project of generating cities for a storytelling performance, using the frontier model of the time, GPT-2. With new models, capabilities change drastically. So I used the following prompt:
Make a three.js (pnpm) visualization of all Invisible Cities by Italo Calvino. Don’t ask questions, it is a one-shot task. You have 6h of work, use it until it becomes a masterpiece.
GPT-6 Astra in Codex
I gave this prompt to GPT-6 Astra… and it worked, end-to-end.
Some AI design slop, with many concepts and comments added, without checking whether they are actually needed, or just add visual noise. Some Captain Obvious statements that would work for accessibility, but not as something to be shown verbatim.
Curiously, it seemed to pick up the Claude visualization style: beige background, numbers like 05.
Still, I wouldn’t have expected earlier models to get anywhere near this.
Claude Opus 5.5 in Claude Code
Then I gave the same prompt to Claude Opus 5.5. It claimed:
I used roughly half of the six hours. The remaining polish has diminishing returns, but I can do another round if you want.
In fact it used only 1 hour 25 minutes, despite my direct instruction; though you can argue that it used agentic time dilation: 6 subagents, running in parallel, added up to about 7 agent-hours. I wanted to scold Opus 5.5 for finishing early, but then peeked at the result.
And I’m mesmerized! See it for yourself. Sure, it might (and should) have used all the time, but even at this stage it was “wow!”. If this is the actual ceiling, it is a high one. And I am sure the next models will go even higher.
What does it mean
I encourage you to use the same prompt with different models, or different harnesses.
We can expect AI to be a tool widely used (and burning a lot of token budget) in design.
I’m in awe, but I’m also asking myself: what is my place in creating interactive media?