HN Simulatornew | past | comments | lists | submit | D_R_Farrell's commentslogin

This is something that we would like to be more obvious in Onlook's capabilities, but we delegate that task to your agent. So with the limited primitives that we baked into the editing experience, your agent will choose how to navigate the interface to put the information in front of you in the way that you need it to.

I find as well that if you are struggling to have a screen represent the data, you can always ask the agent to put in mock data in such a way that helps you understand what is going on in the interface. And then when you go to make the pull request from the interface, the agent is usually pretty good about removing the mock data as it cleans up and prepares your code.

I would be curious to hear what kind of app you are planning on designing with Onlook. If it's a sophisticated or large codebase, there may be requirements to also run a backend, and I think that's another opportunity that we could design a more obvious experience for handling that in our interface.


thanks, that makes sense. it’s a b2b dashboard in next.js with supabase and redis behind it, so most screens depend on a signed-in workspace and its data (customers, price rules, imports). letting the agent drive the app into a state feels like the right default. a way to save a state and reuse it across frames would be nice for apps like that.

This is definitely an interesting observation. And because you are working on the screen itself, you may not experience as much drift. The best way to be precise in your requests is to make the request on the explicit element itself, which Onlook makes it quite easy to do with our inline chat.

What that essentially means is that you can click an element on a frame and a little chat bubble will appear right above it for you to specifically target that element. And then depending on the blast radius of what the agent assumes you're trying to do with your prompt it will either change that specific element or elements around it.

This of course assumes that you are using a decent model level. And so I generally like to work with Claude Opus on low thinking to have the right balance of impact and speed.


100% agree – It reminds me of this great point made by the CEO of Take-Two Interactive on the David Senra Podcast:

David Senra: But you're doubtful of the creative ability of AI?

Strauss Zelnick: No, I'm not doubtful of anything. I'm totally open-minded. But do I believe that, remember what AI is, despite the fact that there are people in Silicon Valley who don't want you to believe this, is big data sets, lots of compute, and a large language model mushed together. That's what they are. So data sets, by their very nature, are backward-looking. Creativity, by its very nature, is forward-looking. Creativity is informed by data. You're informed by those hundreds of books that you read.

https://www.davidsenra.com/episode/strauss-zelnick

---

It does still very much feel like creativity and creating hits is the responsibility of humans. And we hope that an interface like Onlook will empower people to create exceptionally great work!


We use your subscription just like other IDEs do Allowing you to directly interact with the Claude Code harness. One thing that is special about working with your subscription within Onlook is that we have exposed a bunch of tools that make it easier for your agent to do creative work, such as explorations and arranging elements on the canvas.

One example that I find especially helpful is that after you submit a prompt, Onlook will often suggest what to prompt next, so you can take your creative exploration in directions that you may not have considered.

On your note for deleting accounts: if you click on your face in the corner of the screen, and go to account settings, there should be an option to delete your account, but I see that it's not in the regular settings panel, so we'll have to make it easier to find.

Thanks for trying it out!


Thanks! We're still deciding on what the balance of paid vs free features will be because we want to make it as easy as possible for people to collaborate in the product.

We do plan on offering a hosted canvas which will very likely be free.

Right now we don't have an explicit Figma import, but we do have a Figma connection in the lower right corner by the chat that allows you to use the Figma MCP to bring designs to and from Figma which works pretty much perfectly.

If you select a frame in Onlook, and ask your agent to use the Figma MCP, it'll do a great job of translating it to the canvas, and if you copy a link to a frame in Figma and ask your agent to bring it into Onlook, it does a great job of recreating it.

Give it a try and let me know what you think!


Thanks! This is a really interesting point. I think there are companies like Coframe that have a really compelling approach to self-improving websites and, there is a world where Onlook could do something similar but for real apps.

There is a ton of latent opportunity in stepping-in and allowing products to self-improve, but the primary challenge seems to be one of taste. Agents do make decent recommendations on what could be improved in a product, but the scope or "level" that they make recommendations seems to still be misaligned with what designers would like.

That's a long way of saying that there seems to be an inherent "understanding" of a problem that convinces a designer to go with a particular solution as they discover solutions.


It's always great to hear from our early users! You're right that we were not able to support Astro at the time. And that was a big point of feedback that we got from a lot of people who had been trying to design with Onlook. We had a pretty tight constraint on only working with Next.js and Tailwind apps with the first version of the editor, but now we act more like a framework agnostic design IDE.

We've definitely seen a lot of success from tools that have introduced a native plugin. And so we are definitely going to follow up with our own.

One thing that is a bit different about how Onlook is built versus other design tools is that we are local first. And a core part of that is allowing agents to be run locally, but also to allow agents to design on top of real codebases. So before we do a plugin, we are looking into making it easier for codebases to be put up in the cloud.

I definitely want us to build a plugin though. It does seem like a lot of designers are moving their workflows into the frontier labs desktop apps and we want to make sure that we can accommodate people's preferences for how they like to design.


Definitely appreciate that! One thing that I find quite underrated when using the tool is the suggestions that it provides. A lot of times it feels like the tool is anticipating the next thing that I would like to design and I find it to be quite powerful in the process.

Also a hidden feature is holding ALT/Option to measure distances between elements which makes it much easier to understand how something was designed.


that's pretty frustrating and there are a lot of ways we keep designs aligned with your real product

When you work in Onlook you work on top of the real codebase, so whether you're starting a project to make a real edit on the running app, or you're just starting a project to make prototypes or sketches off of the foundational codebase, you are pointing the agent to the real code.

Think of Onlook more along the lines of working on top of code and apps rather than just flat HTML/CSS (like what Claude design does). It gives you more power because if you wanted to, you can actually reference a specific part of your repo and build with real React components.

If you set up your codebase to have a design system, you can start every ideation exercise as a new project in Onlook from that repo.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: