Having dealt with similar problems on the implementation side it's incredible how in this day and age it's still so difficult to pass state through sign-in or sign-up.
Half of the reason appears to be how the Oauth 2.0 standard is written - it purposefully forbids browser-specific items like the hash fragment in the URL. For passing state you have the "state" field but, despite the name, its original purpose was to prevent XSRF attacks, not store data.
Vendor libraries don't go to huge lengths to help here, as they settle on complying with the standard.
What if you're redirected to a different origin? Not a common use case, but it does occur when you e.g. have several services and select where to go in the sign-in form.
Ideally one should be able to pass structured state throughout, but all we have is `state` really.
True, but it's extremely effective at getting registrations, almost to the point that you're dooming yourself without it. Anyone and everyone uses it, big and small. Google can afford any random user to run a few queries before dangling the hook, but it's all the same.
Ux on mobile is very bad. First you lure me with a prompt, but when I type it it's lost and I'm redirected to login page. Ok, I signed up, logged in. On mobile the progress panel is 1cm wide, I didn't notice and I set the same prompt now I have 2 generating and no way to cancel. You (correctly) started title of this post with the word "generate" that is your core feature, everything else I see on mobile is unrelated bullshit, including the carousel, including the "choose your role" dialog. On mobile there should be prompt (50% of screen real estate), a generate button, output/progress (50% of screen real estate) and hamburger button that hides everything else.
The generation haven't finished yet, I will report when it finishes to see if core feature is even usable.
Mobile is tricky, since such design software are very difficult to use on a small mobile screen. We are looking into it though. How was the generation?
Update: on second try the generation succeeded, the output was hallucination, completely unrelated to the prompt. Useless. Image comparing what I wanted and what was generated: https://files.catbox.moe/kplzpk.png
I've been doing a lot of 3D design in Codex GPT 5.5 (I found Opus 4.7 wasn't as good - haven't experimented much with 4.8 or Fable).
OpenSCAD is a parametric CAD programming language, and the models know it well.
The biggest challenge is communicating words like "inside" and "above" to the model - inevitably it's idea of which direction is which is often different.
I can't say I've done anything very hard, but for things like ESP32 cases, or parametric rod connectors it is great.
You can do things like "add snap connectors" and it'll do a great job.
Examples guys. You are telling me a lot and not showing me. Get rid of the home page animation that does nothing and replace it with examples that show real world use cases, step by step.
Would be curious to hear from folks who do modelling:
My first reaction from seeing the examples on the page, is that they are somewhat simple, almost standardized, or come from templates. They also don't look very time-consuming to model yourself. The product advertises itself on the parameterization aspect, but my hunch is that practitioners have their personal library of models which can be rescaled without much work. Is that somewhat accurate?
Unless I know I’m making a ton of variations of a model I’ll usually start from scratch every time. Maybe I’m blinded by my own experience but describing unique geometry with just words seems so much harder than just using regular cad. I design fixtures on a manufacturing line for a living and so I spend most of my time turning people’s ideas into 3d models and nobody can explain what they need with just words. People definitely try to explain what’s in their head but words aren’t enough for anything more complex than the examples on the page above. I always need to see a physical demonstration in person or a napkin sketch.
Tried two very open ended prompts: "Home robot" and "Jet engine". First one returned only two plates connected with 4 rods, second model did not produce any 3d model at all.
Curious whether some hierarchical modeling could be applied here.
Thank you for the feedback! we will consider them both for the website and the product. We have several examples in the product, we will push them to be present on the webpage
I got enough feel from scrolling down. But only just. They need more demoage for sure. Good demo for AI IMO is do 50 canned prompts with cached answers running at simulated speed. Zero AI cost to demo and can be made available without login.
Typed out an entire prompt on the homepage, hit enter, was redirected to sign-up & completed it.
My few hundred word prompt wasn't saved/persisted. :'(
Same. Hated it so much.
Having dealt with similar problems on the implementation side it's incredible how in this day and age it's still so difficult to pass state through sign-in or sign-up.
Half of the reason appears to be how the Oauth 2.0 standard is written - it purposefully forbids browser-specific items like the hash fragment in the URL. For passing state you have the "state" field but, despite the name, its original purpose was to prevent XSRF attacks, not store data.
Vendor libraries don't go to huge lengths to help here, as they settle on complying with the standard.
Why not just use OPFS?
What if you're redirected to a different origin? Not a common use case, but it does occur when you e.g. have several services and select where to go in the sign-in form.
Ideally one should be able to pass structured state throughout, but all we have is `state` really.
Sure but the idea is that your content should be on the same origin after the redirect.
Requiring login after getting user input is a dark pattern.
This bad implementation of it makes it even worse.
I hate it with a passion and I will boycott services that do those inconsiderate things to me.
True, but it's extremely effective at getting registrations, almost to the point that you're dooming yourself without it. Anyone and everyone uses it, big and small. Google can afford any random user to run a few queries before dangling the hook, but it's all the same.
Most dark patterns are indeed done for the benefit of the one who is doing them.
We just updated the product! now your initial prompt is passed by after you sign-up
But you still let the user put in the effort before revealing they'll have to sign up. Just be upfront.
Ux on mobile is very bad. First you lure me with a prompt, but when I type it it's lost and I'm redirected to login page. Ok, I signed up, logged in. On mobile the progress panel is 1cm wide, I didn't notice and I set the same prompt now I have 2 generating and no way to cancel. You (correctly) started title of this post with the word "generate" that is your core feature, everything else I see on mobile is unrelated bullshit, including the carousel, including the "choose your role" dialog. On mobile there should be prompt (50% of screen real estate), a generate button, output/progress (50% of screen real estate) and hamburger button that hides everything else.
The generation haven't finished yet, I will report when it finishes to see if core feature is even usable.
Mobile is tricky, since such design software are very difficult to use on a small mobile screen. We are looking into it though. How was the generation?
It's not tricky. I literally told you how it should look like on mobile.
Still generating, 53 minutes.
As a counter point, I would not like your proposed design.
Why would you use 50% of the screen space when you can have 100% and flip between prompt and render?
Whilst a valid complaint, I'd not even entertain a response with this attitude.
Update: on second try the generation succeeded, the output was hallucination, completely unrelated to the prompt. Useless. Image comparing what I wanted and what was generated: https://files.catbox.moe/kplzpk.png
I've been doing a lot of 3D design in Codex GPT 5.5 (I found Opus 4.7 wasn't as good - haven't experimented much with 4.8 or Fable).
OpenSCAD is a parametric CAD programming language, and the models know it well.
The biggest challenge is communicating words like "inside" and "above" to the model - inevitably it's idea of which direction is which is often different.
I can't say I've done anything very hard, but for things like ESP32 cases, or parametric rod connectors it is great.
You can do things like "add snap connectors" and it'll do a great job.
Can you share prompts and the output? I’ve tried building fairly simple models and it doesn’t grasp that I am building a solid object.
It's not one prompt, but here is a parametric rod connector:
..
..
..
etc etc
Blog, youtube channel and landing page is kinda sparse when it comes to examples.
And I'm not going to sign up just to test it.
I agree with the other ones, have 20 cached examples or at least some videos that show the process from start to finish.
Examples guys. You are telling me a lot and not showing me. Get rid of the home page animation that does nothing and replace it with examples that show real world use cases, step by step.
Would be curious to hear from folks who do modelling:
My first reaction from seeing the examples on the page, is that they are somewhat simple, almost standardized, or come from templates. They also don't look very time-consuming to model yourself. The product advertises itself on the parameterization aspect, but my hunch is that practitioners have their personal library of models which can be rescaled without much work. Is that somewhat accurate?
Unless I know I’m making a ton of variations of a model I’ll usually start from scratch every time. Maybe I’m blinded by my own experience but describing unique geometry with just words seems so much harder than just using regular cad. I design fixtures on a manufacturing line for a living and so I spend most of my time turning people’s ideas into 3d models and nobody can explain what they need with just words. People definitely try to explain what’s in their head but words aren’t enough for anything more complex than the examples on the page above. I always need to see a physical demonstration in person or a napkin sketch.
I had a lot of fun with https://github.com/tomasf/Cadova
A library/toolkit for the Swift programming language to built parametric 3d objects.
Give Claude access to the docs and examples and you can do amazing things.
Tried two very open ended prompts: "Home robot" and "Jet engine". First one returned only two plates connected with 4 rods, second model did not produce any 3d model at all.
Curious whether some hierarchical modeling could be applied here.
Thank you for the feedback! we will consider them both for the website and the product. We have several examples in the product, we will push them to be present on the webpage
No you won't. This is just corporate soothing.
Show us the results. Don't talk about the future. Everyone can yield empty promises "in the future". That means nothing.
Asked for a specific tool and got a tool holder? I hope this project succeeds but it's not ready for prime time HN.
It wants me to log in, so I am their product.
No thanks.
asks you to enter a description, and when entered, hits you with a sign in dickover.
It actually works pretty well.
Great! what did you design?
Not really.
What was your prompt and what did you design.
I’ve tried Gemini (paid) in the past with openscad and then designed the same object myself and I was 10 times faster
Larp.
Like it.
Not signing up if there isn't even a hero animation showing what it does.
I got enough feel from scrolling down. But only just. They need more demoage for sure. Good demo for AI IMO is do 50 canned prompts with cached answers running at simulated speed. Zero AI cost to demo and can be made available without login.
amazing! what are you designing?
https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcQrAPPi...
Account created 7 days ago, post that adds nothing to the conversation. Flagged.