Adrian | JavaScript Mastery

NEW VIDEO: How to Actually Build Mobile Apps with AI in 2026 | A Complete Beginner's Tutorial 🔥 ➡️ jsm.dev/twitter-yt-lin… In this video, you're going to build a full Duolingo clone in React Native, using AI the whole way through. You won't write a single line of manual code, but you also won't yolo prompts and hope it works. I'm calling it 'Practical Vibe Coding': → App framework powered by React Native & @expo → Social authentication & user management with @clerk → Talk out loud to a real-time AI teacher with @getstream_io → Product analytics and tracking with @posthog → Every AI-generated pull request reviewed with @coderabbitai → Styled with @nativewindcss → State management handled by Zustand The app isn't the point, the workflow is. Use this on anything you build next. Grab your coffee. Let's go. ➡️ jsm.dev/twitter-yt-lin…

GLSL used to be the wall in creative coding. I spent so much time learning it so I was never able to ship the effect. Just yesterday I described the idea to @GoogleAIStudio and after a couple of hours, I was able to get a swirl, refraction, blur, all interactive with @threejs Live: swirl-shader.vercel.app

I’ve been experimenting with something I’m calling Practical Vibe Coding. The idea is simple. Vibe coding is powerful, but without direction it quickly becomes chaos. On the other hand, having giant master prompts, 5–10 markdown context files, and endlessly micromanaging AI agents also starts feeling less like building products and more like managing the AI itself. So I’ve been trying a middle ground with: - Single strong architectural context - Visual references - Tiny feature-by-feature prompts Modern models are already surprisingly good at understanding UI, visuals, app flow, and large context windows. Which makes me think the future may not be about creating more complex prompting systems. It may actually be better architecture, better constraints, and faster iteration loops. Basically, giving speed a direction. Been testing this while building a full AI-powered React Native app lately and it’s been working far better than I expected. Full video on it coming tomorrow on YouTube.

Next video! Subscribe and hit that bell on YouTube 👉 jsm.dev/youtube-twitter

@FlorinPop17 Are you scraping me? jkjk :D

Cooking something HUGE! 🤩

The craziest part about modern AI coding isn’t “generate code”. It’s instead a focus on product thinking, UX and architecture …and let AI handle most repetitive implementation work.

Built this hero scene without opening blender once. Entire thing is line meshes only: laptop, cup, notebook, pen, phone, data trails. every line you see is geometry, without textures or shaders @threejs R3F @greensock @GoogleAIStudio - that's the stack line-mesh-demo.vercel.app

Projects that took weeks now take days — but the real shift is how we learn. Adrian Hajdin (@jsmasterypro) shares how AI became a teaching layer in his workflow and content watched by millions. Hope to see you at React Summit!

A year ago, implementing auth in a React Native app meant recording a 30–40 min tutorial And today? I wrote one structured prompt and the whole auth flow was done in minutes 🤯 Sign up, sign in, verification modal, redirects, @clerk integration flow, all generated in one shot and properly working too. AI isn’t replacing development. In fact, it’s compressing implementation time dramatically.

"Agentic coding is a trap." Lars Faye's article making rounds today is worth your time: larsfaye.com/articles/agent… His core point: supervising AI requires the exact skills AI usage erodes. Anthropic's own research backs it. 47% drop in debugging skills. Juniors can't build foundations from reviewing code alone. Here's the weird thing. His prescribed antidote is basically the workflow we shipped in "How Senior Engineers Actually Build With AI": → Specs first, written by you → Planning mode before code → Pseudo-code to close the gap between intent and output → Never generate more than you can review in a sitting → MCP servers (Context7, Figma, Playwright) so AI verifies its own work The trap isn't AI in your workflow. The trap is letting AI replace your thinking instead of augmenting it. Vibe coding atrophies you. Spec-driven orchestration doesn't.

Physics is not hard anymore. I built a physics demo without writing any physics code. @GoogleAIStudio + matter.js did the work. Click around and enjoy the mess. Live: voxel-2d-physics.vercel.app

Showing personality through graphics used to take a studio, a 3D artist, and weeks of work. Now it takes an afternoon and the right references. Live: y2k-dev-hero-section.vercel.app inspo: pinterest Original Creator: @designgemstudio

One thing I’ve realized while vibe coding with AI. The quality of the result depends less on “big prompts” and more on architecture, constraints, context & consistency. I’ve been using a 487 line long AGENTS.md workflow + tiny prompts to build a Expo app feature by feature. And it’s working way better than expected.

it’s wild that I used to need blender + hours of tweaking just to get a halfway decent 3D render now i just... prompted it. and it looks like this. built this with @threejs + @GoogleAIStudio for a trading platform demo. no blender. no 3D modeling skills. something has genuinely shifted in what a solo dev can ship

Been experimenting with a different way of coding lately. Instead of writing huge prompts grill skill me, I created a proper AGENTS.md architecture + small couple of lines prompts for AI agents and tested it on a full Duolingo-inspired app feature by feature using tiny prompts. And honestly… it feels surprisingly close to output you get using huge prompts context md files and such. Going to teach this whole workflow in an upcoming video 👀

If you've been wanting to learn @tan_stack Start - this is your last weekend to grab 20% off our PRO course → jsm.dev/twitter-tansta… 🚀

😎The full React Summit lineup is officially out! 50+ top React engineers. Deep dives into AI, React architecture, performance, tooling, frameworks, design systems, DX & more. Join thousands of developers in Amsterdam & online.

me before AI: "let me read 400 pages of docs to find this one method" me now: "claude figure it out"