Leo ♦️

We built an AI document processing system for Deutsche Verrechnungsstelle AG — one of Germany's largest billing/factoring companies. 2M+ documents processed in the first month. Fully GDPR-compliant. Two-person team from Berlin. Full case study:

Repo is public. Fork it, swap in your own PixelLab assets, and you have your own portfolio world: github.com/duozokker/leon… Stack: Kaboom.js + Vite + TypeScript Tiled for the map editor PixelLab for AI-generated pixel art assets Gemini CLI / Cursor for spatial coding

I built my portfolio as a playable pixel RPG. No template. No Framer. You walk around a retro town, talk to NPCs, and discover my work through exploration. leonaderi.com Here's how I made it and what I learned. The whole thing runs on Kaboom.js, a lightweight JavaScript game library. About 800 lines of TypeScript handle sprites, physics, collision, camera, dialog system, everything. One Vite project. No React. Just a canvas and game logic. Every building on the map links to a real part of my portfolio. LinkedIn, GitHub, my company Artesiana. You walk up, press Space, and a Pokemon-style dialog pops up with info and links. There's a recruiter NPC who has an easter egg if you talk to him three times. The map is a 66x42 tile grid with Wang tile autotiling for smooth terrain transitions between water, grass, dirt, and brick. I wrote a greedy meshing algorithm that batches water collision boxes into horizontal strips instead of spawning thousands of individual tile colliders. That's what makes it run smooth on mobile. All the pixel art was generated with PixelLab. Characters, buildings, trees, tilesets, walk animations. I gave it reference styles and it produced consistent spritesheets across the board. If you do any kind of indie game dev, check it out. Now the interesting part. I coded most of this with Gemini 3.1 Pro inside Gemini CLI. I tried basically everything else. Claude, GPT-5.4. None of them could properly reason about spatial game layouts. When I said "place the bridge at tile 26,22 connecting the island to the mainland" they fumbled the coordinates, broke collision boxes, or just didn't understand what the result would look like visually. Gemini 3.1 Pro actually gets it. It has real spatial intuition. I described what I wanted and it placed objects correctly, understood depth sorting, and wrote clean Kaboom.js code that rendered the way I had it in my head. For this kind of visual-spatial coding, nothing else came close. That said, Gemini is also kind of unhinged. It pushed my API keys to the repo on first deploy. It kept changing things that were already working. Very confident, very wrong sometimes. It has this energy where it "knows better" and silently rewrites stuff you didn't ask it to touch. Setting it up through Vertex was painful. Gemini CLI itself is super buggy, performance is inconsistent, the whole DX is rough. I eventually moved to Cursor with Gemini 3.1 Pro as the backend model. That's what actually tamed it. Cursor gave it the guardrails it needed. Same spatial intelligence, way less chaos. After the core was done, I cleaned up the codebase with GPT-5.4. Refactoring, strict TypeScript, architecture cleanup. Each model where it's strongest. Some more details on what's in there: Mobile support with a virtual D-pad and action button i18n toggle that hot-swaps all NPC dialog and UI between English and German Camera with lerp smoothing, zoom controls, and boundary clamping so you never see black borders Y-sorting so the player walks behind trees and in front of paths naturally Typewriter text animation with skip-on-tap, just like Pokemon Deployed on Vercel, zero config From idea to live: a few days. Repo link in the reply.

This Claude Code skill that clones any website in one prompt is actually wild. No more screenshot guessing, no more fighting with fonts and spacing. It opens Chrome MCP, pulls real assets and code, then spins up parallel agents to rebuild everything properly. One could build serious production agents with this idea: • Use Opus 4.6 as orchestrator to deeply learn the full vision + docs first (Context7 + websearch) • Spawn parallel subagents with proper research phase + tests • Let Gemini 3.1 in Cursor do the frontend magic (it still finds UI bugs no one else sees) • Finish reliably with GPT-5.4 + Codex autonomous review I'm already running something very similar for my self-hosted OCR agents on Runpod after Mistral doubled the price. The difference in speed and quality is ridiculous. What's the best "one-prompt" agent skill you've seen or built lately?

Voxtral just dropped and it actually sounds very promising. Low latency, better emotional expressiveness, natural prosody and it seems to handle long-form content much smoother than previous TTS models. This could be a big step for voice-enabled agents. One could build really nice experiences like: • Reading out processed documents or reports • Voice-first internal tools • Natural audio summaries from enterprise data I'm especially curious how well it integrates with agent workflows (planning → action → spoken output). Would you use something like Voxtral in production voice features, or are you waiting for more benchmarks first?

sauce x.com/_andresjasso/s…

This AI-native UI tool (Tenor) + Claude Code is a big deal. Skip Figma, build real components directly, send them into your codebase. Regenerate from screenshots. Best design-to-code workflow you're using right now?

This is exactly why single-model workflows are getting outdated. Guy rebuilds an entire app from a video in under a minute using Grok for requirements + Claude Code in Plan Mode. Insane speed. One could build serious enterprise agents with this approach: - Use Opus 4.6 as orchestrator to deeply learn the vision + docs first - Spawn parallel subagents with proper research + tests - Let Gemini handle the frontend glow-up - Finish reliably with GPT-5.4 + Codex review I’m already doing something very similar for my self-hosted OCR agents on Runpod. The multiplier when models play to their strengths is ridiculous. What’s the craziest “one-prompt rebuild” you’ve seen lately?

@Devinbuild Codex

Which tool actually helps you? - Claude Code - Cursor - OpenClaw - VScode

@TeeDevh too much overhead Gotta get things done and that (whilst fun) is just another distraction

Why do people still use Vercel? You can just buy a $5 VPS on Hetzner and run everything yourself. More control. Lower cost. More responsibility 🤔

@arlanr same setup lol

my engineer is running openclaw on his raspberry pi 💀

@TusharKapil003 leonaderi.com

Drop your best website. I’m gonna rate it.

@thesayannayak leonaderi.com

What problem does your product solve? Drop the link.

@NoahKingJr build things till they’re done

TELL ME SOMETHING YOU CAN DO THAT CLAUDE CANNOT

@hthieblot @nichtleoo (me lol) and @Luuuca78 are getting rid of bureaucracy right now by building MandantLink (automates german tax lawyers) powered by our own framework Agent-2 (fastapi for openclaw style agents) We processed over 2 million documents so far! x.com/nichtleoo/stat…

Drop or tag an underrated builder/founder who deserves way more attention. If it’s you, self-tag, no shame.

Humans can see in high-res, high-FPS in real-time. Why can't VLMs? Introducing AutoGaze: ViTs/VLMs "gaze" only at key video regions! Up to 4-100x token savings, 19x speedup, and enables scaling to 4K-res 1K-frame videos. 📄 arxiv.org/abs/2603.12254 🌐 autogaze.github.io 🤗 huggingface.co/collections/bf… (1/n)🧵

@thekitze honestly the timing is suspicious. every time claude ships something big openai drops a reorg or a new model announcement within 48 hours. at this point im convinced they have a slack alert for anthropic blog posts. curious which agent setup you're running day to day?

claude code is shipping so hard right now that openai started restructuring 💀

@enunomaduro the chunk size is everything. too granular and you waste time prompting. too vague and it hallucinates. we landed on 3 to 5 step plans before each session, changed everything. do you prescribe the plan or let the model break it down?

your experience with claude code (or any AI model really) changes a lot when you build things step by step instead of asking for everything at once.. this is where most people struggle when they say AI doesn’t work well for them.. build things incrementally and guide the agent: that’s basically what real software engineering is turning into..

I’ve consumed so much AI-generated text that I’ve started writing like one The slop is contagious —

@hthieblot remove bureaucracy. placeholder

pitch me your company in 3 word.

@TJenkinsTampa @factorydoge69 monki brain dont like rejection

@nichtleoo @factorydoge69 Why is it annoying if it is literally correcting you and saving you hours of debugging down the line?

claude confirm all my pre existing cognitive biases and amplify them by sycophantically agreeing with all my ideas. make no mistakes.