John Wayne

Start-up in the day and indie dev at night.
Shipswift | Trovy | Fullpack | 烟火小店 | Pasting

[John Wayne @DevJohnWayne]

Now your pile of hoard can fly.

[John Wayne @DevJohnWayne]

Recreate Airbnb’s homepage and listing pages with this skill.
https://t.co/XQ5aJBfG7D

[John Wayne @DevJohnWayne]

@jacksunny18 Fair point. Each model runs in its own vendor's agent on purpose, so this does measure the pair. Both had the same simulator and shell though. More runs at different reasoning levels are coming.

[John Wayne @DevJohnWayne]

Opus 5.5 vs GPT-6 Astra on 8 real app-building tasks, each in its own coding agent.

Quality: 94.5 vs 88.7
Time: 3h 6m vs 1h 19m
Cost: $52 vs $23

The gap is in how rebuilt iOS screens move. Open source: https://t.co/Cz5F7van19 https://t.co/RN3TMPLc9Y

[John Wayne @DevJohnWayne]

@harden_369741 @Vikramaditya_V1 Thanks, that's exactly why I keep the three scores separate.

[John Wayne @DevJohnWayne]

@blud872 Good guess, that's where a lot of Opus's time went.

[John Wayne @DevJohnWayne]

@youknowwh0ami Ha, pretty much. It really depends on the job.

[John Wayne @DevJohnWayne]

@Xai2Xsi Makes sense. Frontend was the widest gap, but Opus also scored higher on code quality in 4 of the 8 tasks, so it depends on what you're shipping.

[John Wayne @DevJohnWayne]

@LeeWyattCorp Mostly the two iOS screen rebuilds: 92.5 vs 77.2, widest on the stats screen. Both passed every hidden test, so the rest came down to code quality.

[John Wayne @DevJohnWayne]

@Vikramaditya_V1 Thanks! Tests at different reasoning levels are on the way.

[John Wayne @DevJohnWayne]

@GrutzDevid Good catch. Astra recorded its build too, but mostly checked stills and colors, while Opus compared the motion frame by frame. Neither was told to. Runs at different reasoning levels are coming.

[John Wayne @DevJohnWayne]

@not_sacke 100%

[John Wayne @DevJohnWayne]

A battery bug in Trovy's sticker pile: once the first pile appeared, CoreMotion ran at 60 Hz and never stopped, even in the background.

The fix is a counter. No pile on screen, stop. App in the background, stop. Reduce Motion on, never start.

https://t.co/ZHzuGO4lXS

[John Wayne @DevJohnWayne]

Two layers of simplex noise mapped across up to 10 colors. One Metal colorEffect you can drop behind any SwiftUI view.

Open source: https://t.co/ZAUOHhZkvo

[John Wayne @DevJohnWayne]

烟火小店的地图上有全国 4000 多家店，pin 一多就糊成一片。

现在的做法：可见区按 10×10 网格分桶，每桶只留优先级最高的一家，角标写桶里有几家；点一下放大 3.5 倍，桶自然裂开。推荐者头像统一降采样到 120px，MapKit 反复重建标注也不会重复下载。

https://t.co/wMbMdq2Ogj

[John Wayne @DevJohnWayne]

Your agent can't eyeball these from a prompt. Give it the source.

85 free SwiftUI skills, open source, one command:
npx skills add signerlabs/shipswift-skills

https://t.co/SO7AUDjAYe

[John Wayne @DevJohnWayne]

Search in Pasting is a SQLite FTS5 trigram index sitting next to SwiftData.

rowid is the last-used time in ms, so results come back newest-first and paging is just rowid < last. 1–2 characters skip the index and scan: about 90 ms on 100k clips.

https://t.co/H9rTfI2bZ8

[John Wayne @DevJohnWayne]

800 dots, one Canvas, 60fps.

Spherical Fibonacci points plus one perspective divide per dot. Drag the slider and the cloud pulls back into a sphere.

Open source: https://t.co/lOYQJfolDQ

[John Wayne @DevJohnWayne]

@gan_simon Mostly simplify. Full-screen piles have no cap: every sticker scales down to fit (35% floor so they stay tappable), and below 60% the traced bodies swap to rectangles. You can't see the difference at that size, and collisions get much cheaper. Room preview cards cap at 100.

[John Wayne @DevJohnWayne]

Every item in Trovy is a sticker with real physics.

SpriteKit bodies traced from each sticker's alpha, gravity follows CoreMotion as you tilt. One tap and they fly onto a Fibonacci sphere, same math as SWDotSphere (open source).

https://t.co/lOYQJfolDQ
https://t.co/ZHzuGO4lXS

[John Wayne @DevJohnWayne]

How Fullpack turns a photo into a sticker.

Vision's foreground instance mask covers the full frame, so the photo splits into two exact layers: the subject stays put, the background breaks into particles.

https://t.co/EV9nlL1zgy

[Jesse Zhang @Nag1ovo]

https://t.co/BkYwTxffi7
huachu-bird-sticker 开源了：249 种现生鸟类 + 1 种 AI Native 种 huachu bird，完全独立重绘，绝不搬运原库一像素，全部放到了 GitHub。
​花 10 亿 Token 让 AI 假装懂鸟类分类学太累了，���们只花了一点点算力，就让它们全部达成了 100% 的审美统一（其中 6 种甚至学会了骑自行车）。这个项目讲清楚了当生物学��像素彻底脱钩后，我们该如何带着认知与审美，生产出真正契合时代价值的东西，找到安身立命之本。