Hacker News Reader

Comment by nico | Hacker News Reader

This is awesome. At the same time I wonder if we are entering an era in which it’s just better to build your own tools

Need a 3D modeler? Use Claude or codex to build the 3D model first, then build the modeler around it as you iterate towards what you want to build

So many times I’ve had to spend hours learning a software, reading docs, doing tutorials, only to later be able to do a small basic idea of what I actually wanted to do in the first place

Now I can just build only the software I need, for the thing I need it for (or, ala REA, maybe either modify or reverse engineer some existing software, to better fit my workflows)

In that sense, it would also be cool to be able to drive KetchUp with an agent. I’m a big fan of what pen.dev has done (although now their tool is not free/open anymore)

View on HN

I have several personal tools I’ve built and maintained.

It’s always more work than it looks. You can get to mostly usable tools quickly if your request isn’t too difficult, but every time I use them I run into another thing that needs to be fixed, changed, clarified, or cleaned up.

It’s fun at first, but over time you realize you’re spending more time building the tools than using the tools to build the thing you originally wanted to work on.

I think personal software is best for highly specific tools or for things you need to use repeatedly for a very long time.

For something like a SketchUp competitor, I wouldn’t want to try to maintain my own. The number of edge cases and features is bigger than you imagine up front, and the vibecoded version that might appear a few hours after your prompt is going to be very far from the version you want to use.

Isn't the whole argument that ai-driven code will reduce the time spent building because AI can do it now? And a SketchUp competitor by defintion will not be personal software just like SketchUp is not.

AI can’t do it. You still have to hold its hand. You still have to sit there and ask it to do things. There is no fire and forget.

For personal projects that don't matter? Sorry but that's your opinion at this point. I respect it but also fully disagree. Also even in that case, isn't it still reducing time and problems we have with pre-AI personal projects like maintainability and feature creep like you mentioned.

It’s just my opinion that you have to ask the LLM to do things? Seems like reality to me tbh.

> For personal projects that don't matter?

My personal projects matter a lot. For some of them I am trusting my private data in the code, so I need to hold it to a high security and privacy standard. For others, I have hundreds or thousands of users (game mods, helper tools for hobbies, etc.), and I don't want to ship low quality to them.

> For personal projects that don't matter?

If it doesn’t matter then do what you want. By definition, it doesn’t matter.

Some of us actually use these tools, so it does matter. You do have to consider things when it matters.

For my personal projects I think it does matter even more. I've got some perculiar and weird quirks that may even go against best practices (that's what makes them personal projects).

AI can get the implementation done just fine but in my experience it needs guidance because it can't guess what features and details I want / don't want.

The difference to "before" is that it's now possible at all for me to get personal projects to a working state but in sum they still require a few hours every now and then supervising.

I was talking about AI developed tooling.

It’s amazing that you can get something usable quickly, but then every time you use the software you’re back in the AI working on your tool.

Using AI to build the tool becomes the project, replacing the project that led you to think about vibecoding a tool.

I'm coming to think that a hybrid of vibe-coding but with an existing substrate is the right way to go.

For instance in the context of this SketchUp clone that might mean using AI to generate an addon for Blender.

For some of my recent projects I'm using the community edition of Odoo to serve as that substrate because it already has an established community of modules to draw from, ten years or so of real world use, and I don't have to burn tokens generating a user account system with permissions or the data models / ORM to interface with them.

This seems to be serving me well, but we'll see. It could be that I end up in the same the tool becoming the project scenario that you describe, just a little bit later than if I had started from a blank slate.

I’ve been using the vibecoded Lightroom clone and it’s 5000 commits and around 20 releases in yet still has countless issues and missing features. I feel like it will get there eventually but the idea that you should just vibe up your own Lightroom rather than using an existing one is absurd with today’s tech. It’s still a massive amount of work even if it is way faster than ever before.

The vibecoded apps all rely on an army of users exploring the program and reporting all the bugs and edge cases. You don’t want to be doing that on your own.

this is something which is starting to annoy me, and the effects can be seen on HN. someone posting a post or comment talking about their new vibecoded project, then a couple other people chime in, all with their own, separate, vibecoded projects, trying to reach the same goal.

it seems like the open-source space is losing some aspects of collaboration. people used to search for already existing projects on some git-hosters, maybe fork them or put effort into the existing thing. now half a dozen people each run their own race using AI, each solution misses 30% of the features and has x-amounts of bugs.

work together ffs :( i mean, 20 people can use AI to vibecode the same project, why does everyone now feel the need to do everything solo. (because they want to be the prolific founder and cool dude who did the most in the project* instead of being one of many contributors, i assume...)

*or rather, their AI.

I've observed the same, so many forks of forks of forks happening right now rather than group efforts. But I think this is coming from the fact that PR review is incredibly costly to the maintainer while generating code is fast. I've submitted PRs to open source projects which claim to welcome contributions but the PRs just never get reviewed (this was an issue pre AI too). If you wanted to build something, constantly fighting an uphill battle to get PRs merged is less appealing than just creating your own project where you can work full speed on whatever you want.

I'm not too sure what the future here looks like and if this is even a problem, but things certainly are changing. These days submitting a well written bug report or feature suggestion, or validating existing reports is more useful than sending a PR.

maybe the focus should shift a bit to using AI for PR reviews. adversarial reviews, summarizing it for a human. we'll see where time takes us...

What is the point though. The maintainer can generate the same code you can. And they can iterate on it quickly without having to go through a review process.

Day jobs and lives away from the PC.

I believe the goal of most of these people is to make money.

Weird that they are all giving the source code away for free and failing on the whole to even ask for donations.

All?

ArtCraft is definitely a play at selling subscriptions.

The slew of "oh yeah, blah blah blah blah blah, that's why I built [X]." comments on every AI post now?

Click through and many you'll find "Pricing".

It's a gold rush.

1. Go to any ArtCraft repo on GitHub

2. Click "Fork"

3. Profit

Classic besides the point.

> work together ffs

Right, and that's something I already argued for since the very beginning of GenAI, namely that suddenly being very costly, economically and ecologically, is not that bad if you build upon what exists and collaborate on progress.

... except that's exactly WHY vibecoded project exist : they do NOT have to give a shit with existing codebase and agree on weird rules. They just start from "scratch". In fact I'm convinced there is the Ikea effect at play, namely that the PM of the vibe coded project is emotionally attached, even proud, of the result they think they provided somehow, despite them not actually doing it. Obviously for-profit GenAI companies have a big incentive to maximize token use and will never promote workflows and tool that minimize the one thing they monetize.

So even though I agree in theory, in practice I think there is little to no overlap.

My experience has been the opposite. I have two vibe coded tools that I use frequently and I spent a few weeks of occasional prompting and now they are basically done. I haven't touched the code for one of them for a month and I use it most days.

To be fair neither is quite as complex as a whole 3D modeller.

I think this is the key:

> To be fair neither is quite as complex as a whole 3D modeller.

Once you cross a certain complexity in your project, it becomes quite work intensive even using AI to do the coding. All the back and forth and cleaning up, exploring ideas and starting from scratch again takes time and effort and maybe most importantly, extensive human thinking. Just to clarify, I am not against AI and use it extensively myself. But vieb-coding all your tools is just not really a feasable idea for most (time/token-cost constraint) if you tools are of a certain complexity.

Actually I think "complexity" is probably not quite the right measure. You can vibe code very complex things, as long as they are in the "easy for AI" zone - clear goals, ideally non-graphically, ideally there's a specification or even better a reference implementation, best an automated way of checking correctness. CAD is about the worst case (and very complicated).

But you could easily vibe code, e.g. an H.265 decoder, even though it is super complex.

Why do you even maintain it if it's for you?

I have lot of tools that work for me.

When externalities break them, it's usually one prompt to fix them.

and when it works on your machine, but fail when exporting on an old macos 14, you only realize when someone tell you...

My experience is it's short to get a working tool and long to finish. Many 3D modulers (blender for example) have literally 1000s of features. Yes, you might not need all of them but you probably need 100s, each with quirks and UI/UX decisions. Just iterating on camera controls can take days. You get an orbit camera working via a prompt. But then find it only zooms in/out so far because the LLM decided that was ideal. So you prompt to fix that. Then you find moving the camera is not zoom aware so it moves in meters per frame and when you're zoomed in things move too faster. So you prompt to fix that. Repeat 10000x. Every feature, extrude, bevel, chamfer, subdivide, select vertices, select faces, select edges, etc etc all require this handholding and finese.

At least that's my experience.

Small tools with few features is great. a 3D modeler is not that.

Yep having the same experience right now with some curve/spline tessellation into triangle strips. There are many degenerate edge cases and when you ask the "ai" to handle one it breaks something else etc. Makes me think that if I had just hand coded it myself I would intimately know the code and get it working faster than prompting. Gah.

> Yep having the same experience right now with some curve/spline tessellation into triangle strips. There are many degenerate edge cases and when you ask the "ai" to handle one it breaks something else etc. Makes me think that if I had just hand coded it myself I would intimately know the code and get it working faster than prompting. Gah.

It gets so much worse. At some point you start hitting edge cases where there is no clear right answer and you have to come up with a kernel wide coherent approach to things like 1ulp differences. Parasolid is full of decades of kludges on top of kludges and giant switch statements trying to preserve backwards compatibility as they explored the space.

LLMs also seem prone to making weird decisions that no human would have done. You can very quickly have a broad implementation - often with features that are great and that you didn't even know you wanted - but when you dig into the features, the weirdness reveals itself bit by bit. If you were developing by hand you would have used a sensible range for zoom from the start, but the LLM doesn't seem to work that way.

I suspect people saying "everyone will vibe code all their software" have never actually tried to vibe code a meaningfully complex application.

I really appreciate this sentiment, and at the same time there are real opportunity costs to building and maintaining your own tools. Even in the era of tokens.

I think it comes down to where you want to spend your time and attention.

https://www.winestockwebdesign.com/Essays/Lisp_Curse.html

I definitely find it easier to create the subset of blender that I need than learning blender, even if it is throwaway.

And a fantastic learning experience

To offer a counterpoint, I’m currently spending a lot of time on the other side of this: having my agent use blender in the background and build on top of it. My agent can drive it well and I’ve benefited from its import capability, renderers, layering, and simulation. I get a lot from blender, even though I don’t ever see its UI.

I ... really..? What sort of stuff are you wanting out of Blender? If it's 3d printing, you probably literally just need to do the soda can or donut tutorial - not even the whole donut tutorial.

Like, I have to imagine that 80% of the "quick stuff" people need to do in Blender would be done with:

* create quad * exude like 5 times * done

Maaaybe

* perform boolean operation between object and a circle or 3

---

Like, I have less than 200 hours in Blender and I was pretty functional after like ... 30 minutes to an hour of a single tutorial.

I don’t see how this could be faster (and not buggier) than generating a script to make what you want in Blender for you.

You’ve likely just switched hobbies to your Blender clone instead of what you wanted to make.

Interesting perspective for me I find it more interesting to learn the transferable skill set behind using this kind of software. Sure I can make a software that does exactly what I want when poked but then I do not learn any transferable skills.

> At the same time I wonder if we are entering an era in which it’s just better to build your own tools

I wonder if Claude Code and other similar tools start to maintain their own base apps that everyone keeps asking for? This would be like a "cache" that makes the requests cheaper/easier to deliver?

Then everyone's "own" apps is really just a customization layer on top of a solid base that already does all the heavy lifting.

So many people are building their own versions of:

- PDF editors

- Word processors

- 3D editors

- Video editors

- Pixel-based image editors

- Vector-based image editors

- CAD editors

etc.

In some of these there are a few flavors as well that are major architectural choices. But most of what people are doing is just playing around with UI and file formats. Much of the functionality is generally the same.

Sigh... are we really just going to re-invent plugins?

We do need a safe layer for non-engineers to customize apps.

All these clones of Photoshop and similar tools are mostly dead ends, they are experiments to see if you can do it, but there will be a consolidation around regularly maintained and powerful and complete apps that likely also offer support.

These experiments are dead-ends for the most part long-term. Fun play things, but they will not gain significantly user bases.

Isn't this more of an argument to use common OSS choices? Like, encourage people to use Blender for 3d modelling; LibreOffice for documents; Gimp, Krita, or one of the many other art programs for ... art; OBS for simple video stuff; and so on and so forth?

> We do need a safe layer for non-engineers to customize apps.

Why? Every time someone cripples software in the name of safety it just gets replaced with more capable software that isn't.

Plugins have long been troubled. It’s hard to make a plugin interface that is actually useful while also not breaking on every app update. And plugins have been such a massive vector for distributing malware.

I guess I'm too used to the collaboration features of word processors like Google Docs. If I build my own, those go out the window otherwise everyone has to sign up for my word processor and learn my word processor's UX.

For the rest, at least so far, I haven't had the collaboration experience but for word processors I can't go back to local files that can't easily be edited/viewed/commented on by multiple people or myself on every device I own.

yes but tons of new features introduced making things interesting

We have existing tools for this. You fork the repo of the closest thing and have AI tune it to exactly what you want.

I'm pretty sure this already happens to some extend. Look at the default website output you get from the different models, there tends to be a 'default' look/approach it takes.

I think they all have been post-trained on specific tasks like building sites, apps, etc and that is a large part of the reason they are pretty good at it.

Honestly i think that now is the chance for established software to go with the flow and I think the major way to go with the flow is create are adapt software to be extensible (so that everyone can let claude or opencode or whatever add their functionality of choice without having to rebuilt the wheel from scratch everytime) and secondly direct mcp access to the software itself.

I think that could be true once you're experienced enough to know what you want.

Before that, though, I think you have to learn what the experts do and what can be done and what you should want. And then you can start to make your own opinions.

By then, unless the major players have done it really wrong, you're not going to make something better than what already exists.

Right now, it's different, because we have that experience prior to the ability to just wish up an app and have it happen.

I literally just had Claude build a 3D modeler slice of Blender that I wanted but that also works in VR [video: (https://www.reddit.com/r/aigamedev/s/q5Y6cskYb9)]

I think the value of adapting full tools to your own workflow far outweighs the maintenance costs nowadays. The tool above for instance took my agent a couple hours and it was good to go. It wasn’t a small slice by any means either: it supports edit/obj mode, extrude, loop cuts, modifiers like subdiv, save/load, undo/redo, bones, auto rig, posing, and more.

Mind you, this is a slice of an open source tool which was still outside of what I wanted from it. Cost wasn’t even part of the equation here.

That's a pretty neat looking program.

Since Blender is OSS, you could probably use one of the existing VR views for Blender and write a bunch of in-app Python scripts to do what you want, if you wanted to also be able to output to Blender files and work within Blender already.

This looks like a pretty well-trodden area in Blender.

Yes I could fiddle and adjust - But see, i dont want that. I want a modeler that works how i want, some things that work like Blender and others that work a different way.

Adding scripts into an existing product does not serve this.

I think we are entering into a new era where we don’t need to accept fully featured software with all of its decisions should they not fit ours.

You might be surprised just how much you can rewrite of Blender's UI if you want to.

And you get battle-tested software and a huge collection of existing plugins, operational modes, export formats, all that. Fully bespoke software for everything you want is a way toward losing interoperability between people and platforms.

It doesnt have to be one or the other: you can have custom software that still works within existing/interoperable formats, as I’m doing here (im not planning to reinvent OBJ/FBX/BLEND)

But building up as opposed to building down or rewriting is freeing.

I can pull in ideas from Houdini and mix them in. I can use Blender’s source as reference for specific implementations should I need. I can keep my tooling specific to the projects Im working on and design it to fit that as opposed to build a workflow within a software meant to be general purpose.

IMO, this kind of full customization is going to become far more commonplace.

Very cool. I’m in the process of creating a video game and building all the tooling around it, both for me and the players. And also, for doing things like recording a quick sequence of gameplay, “sandwiching” it with an intro/otro and posting to YouTube, then ingesting the engagement data to show on the internal player… still early days, not enough data to start optimizing on it, but I’m really enjoying building exactly what I want as I go along

One of the tools (built for me to score options of different assets): https://id-le.com/review/intro-logo-animation

Video: https://youtube.com/shorts/tlI9_XBuqVk

Repo: https://github.com/nicobrenner/id-le

Online: https://id-le.com/

You mention that you have no plans to share. Any chance I could persuade you otherwise? I maintain Open Brush and Open Blocks and I'm very interested in getting something more complex working for 3d modelling and actually shipping it to the various app stores. I've got a few different approaches in mind but using Blender as a base it quite appealing.

I could start from scratch obviously - but seems a shame.

This really isn’t Blender as a base- more like uses Blender as a base reference.

I think both those apps are great and happy to know folks like you are maintaining them; this one though is simply more of a personal software than a path to an OSS project!

With strong building blocks, perhaps. But I don’t want to maintain millions of lines of AI generated CAD kernel (or pay the token and wait time). I’m also happy not having to make design choices in that space considering I have 0 experience with it. It’s still a liability.

> I wonder if we are entering an era in which it’s just better to build your own tools

Not necessarily. But only if what you need is not available in open source or not achievable by forking open source. Otherwise it’s a waste of money (token) and attention. This is a significantly smaller pool that “build [all] your own tools”.

Although, AI mega corpos would love if we do constantly re-build everything. You can tell by the tendency of the agents to write everything from scratch instead of importing the code that does the same thing or better.

Typically what takes time isn't learning a shortcut here or there, rather it's to understand how a proficient user thinks and THUS why the tool work in a "weird" way for newcomers. It's easy to click and paint, why would pro user layers? It's so complicated... then you watch how pro do it, and it's because behind the complicated steps comes a lot of power.

I have no doubt as a newcomer you can make software that will do what you want, but most likely the outcome will no be good, not because of yourself or the tool, but rather because the expertise to map the problem space to the solution space is not trivial, and that's precisely what requires a new way to think, and thus tools that aren't so easy to learn. It's not poor UX.

"So many times I’ve had to spend hours learning a software, reading docs, doing tutorials, only to later be able to do a small basic idea of what I actually wanted to do in the first place

Now I can just build only the software I need, for the thing I need it for"

Or another way to put it for devs “A language that doesn't affect the way you think about programming is not worth knowing.” ― Alan J. Perlis

I also have my own tool. I vibed a browser-based OpenSCAD harness. It’s a simple SPA with a chat loop. It stores everything in the user’s browser and requires the users to bring their own tokens (URL with API Key or OpenRouter with PKCE)

I host it here: https://vibe3d.spoerri.dev - my partner enjoys it because she can bring an idea to print pretty fast.

Maybe a similar idea could be interesting for KetchUp as well.

I've been using SolidCraft with MCP and Codex to model things. Pretty wild. I gave it a photo of a pencil sketch and a paragraph; a few minutes later I had a 3D model. https://solidcraft.io/mcp/

This was only possible because a well thought out, working example could be used as reinforcement learning for the LLM to benchmark it's own development against and improve its output. Had this not already exists in it's current form, I doubt the LLM could have built this on its own.

So, I guess yes, if there is someone's hard work you can copy and steal and make your own, I guess that's true. But try to create something that no one has ever built yet.

I have always been skeptical of the reality of this, especially for businesses. But then one person on our CSM team spent a week building an alternative to Gainsight, saving the company like $90k a year. Granted, it required a technical person to do it even with vibe coding, but it doesn’t take much of his time to maintain it.

Yes, see:

https://news.ycombinator.com/item?id=50406927

For an example. Lots of people are now vibe coding their tools.

Looking at CRM tools these days and they are so complicated that I’m thinking of just using my tool and add features as I need them.

Would love to hear advice from people who use CRM tools on the job.

Better as in you'll be happier with the end product after a dozen iterations and two dozen small tweaks? Sure.

I think people underestimate what a time investment building your own tool that's of at least medium complexity is, though. It might be worth it if it's something you'll use regularly for a long time, but I don't foresee it being the norm. Even assuming no bugs ever or no performance issues ever and all the architectural decisions made were appropriate (pretty far from current reality), merely describing what you want and how it should behave and look like is a long process, and one that isn't all that enjoyable once you go through it once or twice already.

The existing software that has been around for several years and you kinda dislike is probably much more reliable than your own homebrew, too. Doesn't matter in some cases, matters hugely in others. Learning that your custom video gallery organizer just vanished half your family videos probably isn't very enjoyable, but at least it's easy to recognize it happening. Something more abstract like CAD or 3d printing slicer, ehhh...

Skip the 3D modeler and have the agent do the operations directly instead.

A good visualizer might be useful thought! If that part of the work needs to be validated.

Ask any 3D creative and you’ll know that AI is not yet sufficient to generate 3D models at the quality that most industries need them at.

you can try fabivo, its not for all uses, but i think the ai agent is great.

Why are people suddenly entirely unwilling to invest effort into their projects...

Wat is the point of personal projects if in the end you learned nothing

Once upon a time, not so long ago, people gained skills by learning about things instead of outsourcing all the hard parts.

By outsourcing the modeling to Claude or whatever, you don't learn how to build 3d models on your own. You've become dependent on the LLMs, and you're trivially replaceable by someone who can build 3d models on their own...because they can also learn to tell Claude to do all the work.

I mean, there is value to artisan work (paying for it), and there’s value in quick stuff; but I’m not sure LLMs can do either in the case of 3d art.

Their models are horribly topoligized, which will matter for most use cases for models; and if you’re just using it to make a cube, it’d be better to learn the related program to do that.

People tend to ask LLMs to make full art pieces (a decade of specialized knowledge and muscle memory), not “draw me a square”.

In 3d art, there is a gulf between the easy stuff and the hard stuff.

Tl; dr: for 3d art, I agree.

This just sounds so horribly inefficient though i’m sure the makers of LLMs love you.

Reply on news.ycombinator.com