把 Anthropic 的 S-1 链接甩给 Opus，它创作了一座菜园。
角度是它自己挑的。招股书里写着，算力通常要提前 12–18 个月签下，于是算力账单成了棚，先搭好；营收是豆苗，往上生长。
六个季度，豆子从 3.76 亿美元长到 116 亿，涨了 30.7 倍，棚只高了 5.9 倍。2026 年一季度，豆尖越过棚顶；二季度开了花，首次经营盈利 9.74 亿。
它还翻出一个反直觉的数字：同一季度净亏 663 亿，九成是可转债跟着估值上涨做的账面重估。估值越高，账面亏得越多。
读招股书、选角度、抽数核账，agent 按 SOP 完全自治，挑刺我来 😂。
种菜传送门在一楼，园子里还有彩蛋哦 ⬇️
Once the damselfly was done, I wanted to see if that same skill could handle any subject, especially infographics. So I extended it with rules for turning real data into a world you can explore.
The Short Blanket is the first thing it made after that. Honestly, it turned out better than I expected.
How this came together:
It started with a different piece: the life of a damselfly, from egg to egg, as a 75-second hand-colored copperplate film. Every frame drawn in code by Opus.
While building it, I kept writing down what worked (and what kept going wrong) and turned it into a reusable skill.
Try it yourself: drag around, zoom in, tap a bed to see the numbers.
glintiary.com/short-blanket.…
Saw all those frame-by-frame, hand-drawn animations going around my timeline and got curious: what if Opus took a boring data table and turned it into a little stage you can actually walk into?
So I handed it a sleep dataset. It built a moonlit dorm: one bed for each stress level, and every blanket is exactly as long as the hours slept. Drag around, zoom in, tap a bed to see the numbers.
Stress, not steps, decides how long your night is.
Honestly? We're barely scratching the surface of what this model can do.
Link in the reply ⬇️
Fins drawn along the bending body look fine on a shark.
A humpback stays curled all the way round its bubble net, and its flippers are a third of its length. same method, the outer one balloons (koi too).
So fins are pinned stiff at the root now. porting was mostly re-checking what quietly worked.
三个踩出来的坑，留给要上线 3D 页面的人：
手机上打字越打越慢，是帧率掉下来：动画时钟每帧最多推进 1/20 秒，帧率一掉到 20 以下，整段剧情就跟着放慢。
先查到的是纸面贴图：每敲一个字，那张 400 多万像素的贴图都整张重传一次。改成只传落墨的那一小块，快了不少，但还没到位。
剩下的在桌面浏览器上复现不出来，就在页面里挂了块性能面板，直接在手机上看每帧耗时：总耗时远大于主线程，说明卡在显卡。顺着查到右下角那排按钮的毛玻璃，3D 画面每帧都在重绘，按钮底下的模糊也跟着每帧重算。去掉毛玻璃、树影降频，从 8 帧回到 26 帧。
体验 letterfall.glintiary.com
这回从设计到发布，整条流水线都交给了 Claude Code
1️⃣ 按需求搭出可交互、带声效的实时 3D 场景
2️⃣ 对照历次积累的踩坑记录精修
3️⃣ 排一段首尾相接的剧情，导出循环视频
4️⃣ 给社交平台做封面卡片 (ogimage)
5️⃣ 发布到 Cloudflare，得到一个打开就能玩的链接
6️⃣ 维护仓库里各个案例和文档之间的双向链接，下个同类项目直接接上这次的沉淀
能一路顺下来，靠的是这个 LLM wiki 里沉淀的个人画像、方法和脚本，我只管说清楚想要什么。
上次说建模学晚了就不用学了，这次拿打字机试了试能做到多沉浸。
在浏览器里真能打字：每个键是一根连着字锤的杠杆，纸上有纤维、逐字压出的凹痕和背面透出的反字。敲击声是代码现合成的，照一台 1940 年 Erika 5 的录音标定（agent 自发）。中文走模拟拼音，选好字再一个个敲上去。
整台机器没有一个模型文件、一段录音，全是 Opus 5.5 写出来的。⬇️
