AI × 设计对照：同一句提示词，Opus 5.5 和 GPT-6 Astra 各做一个 3D 可视化 AI × design comparison: one prompt, Opus 5.5 and GPT-6 Astra each build a 3D visualization

详细介绍Details

Quesma 的 Piotr Migdał 2026-10-07 发文：用同一句提示词，让 GPT-6 Astra（在 Codex 里）和 Claude Opus 5.5（在 Claude Code 里）一次做完卡尔维诺《看不见的城市》全部城市的 three.js 可视化。作者自己的记录是：Astra 用 53 分钟、约 10 美元，能跑通，但他觉得堆了很多没必要的说明和装饰；Opus 5.5 用 1 小时 25 分、6 个子代理并行、约 74 美元，他看完很惊艳。两个成品和源码都公开，可以自己点开比。适合想知道「现在的模型做交互设计能到什么程度」的设计师，也可以直接拿他的提示词去试别的模型。 Piotr Migdał at Quesma, 2026-10-07: the same prompt asked GPT-6 Astra (in Codex) and Claude Opus 5.5 (in Claude Code) to build, in one shot, a three.js visualization of every city in Italo Calvino's Invisible Cities. His notes: Astra took 53 minutes and about $10 and worked end to end, but he found it full of unnecessary notes and decoration; Opus 5.5 took 1 hour 25 minutes with 6 parallel subagents and about $74, and he was wowed. Both results and their source code are public, so you can compare them yourself. Good for designers who want to see how far models get on interactive design, and the prompt is reusable on other models.

为什么推荐Why we recommend it 提示词原文：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. 用时和花费都是作者自述，本库没有复跑。好不好看是作者的主观判断，只有一个提示词、一次运行，不能当成模型排名。 The prompt, verbatim: 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. Times and costs are the author's own; not re-run here. Which looks better is his opinion, from one prompt and one run, so it is not a model ranking.

上手路径How to start

1. 先跑一次安全范围Run a safe first pass

   先在可撤回、不外发的范围内试跑。本条未批准前：自己复跑时注意花费：作者那次 Opus 5.5 约 74 美元 API 费用。。 Try it first in a reversible, non-outbound scope. Until approved: Mind the cost if you re-run it: the author's Opus 5.5 run cost about $74 in API tokens..

2. 存 SkillSave a Skill

   把跑通的步骤存成 Skill，下次同类任务直接复用，而不是重新口述。 Save the working steps as a Skill so the next similar job reuses them instead of being re-explained.

3. 再上 RoutineThen add a Routine

   确认输出可靠后，再考虑定时或事件触发；例行默认仍应停在待审。 After output is reliable, consider a schedule or event trigger. Routines should still stop at review by default.