I wrote a short spec for the look, ending on a slow push in as the ring crosses the planet's edge. Claude Code built it in Blender and I sent notes each round. v1 was a tan orb with fake sticker reflections. v2 went way too dark. v3 fixed both. 3 rounds to get what i wanted..
Did a desk lamp loop this morning and liked it enough to go again. Wanted something cooler, something inspired by AI and space. A smoked glass planet with a thin glass ring and little comets running through light threads inside. Third round is the one I liked.
@lanre_theade that split makes a lot of sense to me. when your design system is already in there, Claude has rules to follow and it moves fast. a brand new page doesn't have those rules yet, so I get why you'd figure out what it should be in Figma first
@designhorf the check before every commit is the part I think most teams skip. the model can read the rules fine. keeping to them over a long session is where it drifts. curious what your check catches most, tokens or spacing?
@JeffWitters yeah "whatever screen the agent saw last" is painfully accurate. it copies the closest example, mistakes and all. what I keep seeing work is plain text rules the agent can actually read, with the reason written next to each one
Gemini 4 Carbon has been producing a lot of my internal work with surprising efficiency and speed. Personall, I think the reasoning capabilities have significantly improved regarding personalization and agent based conversations, making them feel more tailored and delightful.
@Selenainux yeah this is real. those details used to die in handoff because they were too small to ticket. now I just build the hover myself and tune it until it feels right. the button label one is my favorite though, it costs nothing and it's still real craft
@angeloldesigns@referodesign love that you kept the content the same, so every difference on the page comes from the system. I think the default AI look mostly comes from the model having nothing real to read. which of the three held up best once you went past the hero?
A is the Claude Code page. I scored it 82.4 out of 100, and Claude Code took 10 minutes 0 seconds to generate it.
B is the Codex page. I scored it 81.6 out of 100, and Codex took 2 minutes 28 seconds.
Both pages, the exact prompt I gave the two tools, and these scores and times are on PromptOff, a site where I publish side-by-side design comparisons of coding tools that each get the same prompt. https://t.co/hOhxcKpBoU
I gave two AI coding tools, Claude Code and Codex, the same prompt: design a launch page for a folding desk lamp with a 3D lamp you can drag to turn and take apart. I didn't edit what they made. The video shows both pages.
Which page would you ship, A or B?
@akshanshgusain I'd lean React. the lag on big boards feels like the more fixable problem to me. faking real states and behavior on a canvas is where handoff gets painful. who's it mainly for, designers or design engineers?
@mayank0324 the error red story is so real. a model takes a color name literally and runs with it. that's why I like naming tokens by what they're for, so your brand orange never ends up on a warning
@byleeholmes the logo spinner is such a nice call. loading is where most apps feel the most generic to me, so putting their brand right there makes even the wait feel like theirs. did the pixel font make the motion on it easier or harder to get right?
Figma shipped Code Layers to everyone, and you can now build complex flows with prototype noodles.
code layers inside the prototype canvas means real logic and motion states live next to the frames, not in a separate handoff. the test is whether easing and timing hold up once flows get branchy.
who's building with this yet?
@smasithick loving this. little details like that are what make a whole system feel cared for. my worry with a three finger gesture is people finding it at all. does the app teach it anywhere, or is it just there for whoever stumbles onto it?