Raykin | AI Skill for kids & families @RaykinStudio

The process I'm keeping: Photo + measurements → CAD → simplify for the actual mounting → check fit and removal → round edges → check again → slice. The two-part design has passed the digital fit and slicing checks. Next: print it and see how it feeds.

I needed a holder for a coreless toilet roll, stuck to the wall below our phone shelf. I sent Codex a bathroom photo and two measurements. Codex made the CAD and exported files for Bambu Studio. Then came my revisions: add a removable splash cover, simplify the bottom frame, shorten the side braces, round the edges. @OpenAI @BambulabGlobal One small change had a knock-on effect: rounding the clips meant raising the cover about 1 mm. We also checked the lift-and-pull path for changing rolls.

keyboard_arrow_left Previous keyboard_arrow_right Next

I'm building #K12Play, a world for kids to explore. The first model (GPT-5.6 Sol) gave me a red pitched roof. After more work, it still looked like a wavy sheet. The height and depth were wrong. I isolated the roof and described it as clay: start with an arch, pull some edges down, lift others up. Shape the back, too. We made a 9×7 B-spline surface with adjustable parameters. I checked the CAD views and white model before rebuilding the hall around it. Then came the material problems: obvious repeating tiles, plus UVs lost during mesh batching. I had to describe colors and patterns against the roof's crown, shoulders, and eaves, and check the mapping inside the game. My SOP now: - Set a visual target and list the differences. - Isolate the hardest part; make its shape adjustable. - Review CAD views and the white model before assembly. - Add materials; verify their placement and UVs in-game. - Walk the scene. Pick the biggest visible problem for the next round. Design → CAD → white model → game.

keyboard_arrow_left Previous keyboard_arrow_right Next

That’s the workflow behind the showcase. A target to compare against. A priority for the next round. A review gate before integration. I’m sharing both the things I make for kids and what I learn getting them built. K12Play Universe. Still building.

8/ Use CAD to inspect buildings and spatial relationships. Check dimensions, connections, intersections, and clearances before making the scene more elaborate. CAD gives me structure I can inspect. Then the actual render gets its own visual review. Both checks matter.

I’m building a world for kids with AI. I burned a lot of tokens on details while the big visual problems stayed. What helped: a visual target, a priority order, and a review gate. 8 practical lessons from K12Play’s character and world iterations 🧵

I’d recommend building reusable character models with a human skeleton. Workflow: first use imagegen to generate four orthographic views (axonometric, top, left, and front), then have Codex model from those references and iterate by comparing against the design sheets until it matches. For the interactive part, Godot is more agent-friendly than Three.js—you can scrub the timeline, record the viewport, and work across multiple windows more efficiently. That said, if the end goal is a web deploy, Three.js is still smoother. For pure story/dialogue scenes, skip full 3D modeling; it burns a lot of tokens.

I made an English lesson for my daughter. It turned into a playable 3D scene. Photos → animation → a character you can move around. Codex helped build it. The early anatomy was rough. We kept fixing it.She loves it. Whether the tenses stick is another question 😅 @OpenAIDevs @OpenAI

keyboard_arrow_left Previous keyboard_arrow_right Next

I used AI to build a planet. Then I went for a walk on it. Down to a tiny island. Through the grass. Past golden roofs and shell-shaped huts. A world for kids to explore, still taking shape. K12Play Universe, Day 5. Built with GPT-6 Astra and GPT-6.1 Sol. @OpenAIDevs @OpenAI

This is what great technical visualization should do: make the invisible legible. You’re not just showing the object—you’re revealing the process behind it, which is exactly what I want for invisible events. Beautiful work—would you be open to sharing your workflow? Thank you.

You can try it yourself here. Fly with the birds over the stormy sea, sound on. …eather-study.joewalker95.chatgpt.site/storm/

keyboard_arrow_left Previous keyboard_arrow_right Next

I gave GPT-6 one image and a prompt. Asked for an interactive film. It used Blender and gave me back a single HTML file. Open sea. Storm clouds rolling in. Two birds skimming the waves. The camera moves like a real cinematographer is behind it. And it has sound 🎧. I just sat there watching it. Shocked. Every feather is built and textured. Image 3 is one of them up close. @OpenAIDevs @OpenAI

K12Play Universe, Day 3. New trick: I let AI review AI. I generate a target concept with GPT's image model. Then Codex compares it with the live render and writes an HTML review report. Area by area: what's different, why, and what to fix next. That report sets the focus for the agent's next iteration. Pitfalls I hit before this: 1. The agent forgot to use textures. Lots of tokens burned. Buildings didn't get prettier. 2. Code-built buildings stay simple geometry. They never feel like art. 3. No spatial design. The agent polishes single objects and forgets the overall mood. Images 3–4 are today's result. About 10% better than yesterday 😂 @OpenAIDevs @OpenAI

keyboard_arrow_left Previous keyboard_arrow_right Next