Two Prompts: Addy Osmani Made a three.js Halloween Mini-Game
Addy Osmani used Opus and three.js to make a Halloween mini-game. In the comments, someone asked how many prompts the fog and ghosts took, and he answered that the whole thing took two prompts. Some followed up asking what the second prompt fixed, and others asked whether the model wrote the shaders too; neither got a reply. "Two prompts" sounds like zero barrier to entry, but the public information only gives the result, not the validation process.
Addy Osmani posted a three.js mini-game on Halloween called Trick or Clawd. He wrote in the tweet that it was made with Opus (a Claude-series model) and three.js. The video shows a Halloween-style 3D scene with fog and ghost-like wraiths.
Addy Osmani is an engineer on the Google Chrome team and has long written about web performance. three.js is a JavaScript library for doing 3D in the browser. Putting these two together isn't new. What's new is that reply in the comments.
Someone asked how many prompts the fog and ghost part took. Addy said the whole thing took two prompts. The first finished most of it, and the second helped refine it. In the replies, someone else followed up asking what exactly the second prompt fixed. He didn't expand. Another person asked whether Opus wrote the shaders as well or only the scene. That question also went unanswered in the replies.
So "two prompts" is only a description of the result, not of the process. A model can fill in a lot of code at once, provided the person asking knows what the scene needs, knows the structure of three.js, and knows how to judge what's wrong after generation. The parts not mentioned in public replies are often exactly where the time is spent.
Someone also made a joke with the command line. Running claudie --trick-or-treat gave a terminal error: unknown option '--trick-or-treat'.
If you want to play it yourself, you can open this link: trick-or-clawd.claudepages.dev. Addy gave the link in a reply.
Two prompts are only the result. What's more worth looking at is that the person asking knows what to validate. Once the model takes over large chunks of the task, human judgment becomes more valuable instead.
发布时间: 2026-11-01 13:30