Solo developer rebuilds vibe codes buggy Adobe Creative Suite "alternative" in Rust using Claude
I am not going to deny it is pretty impressive that current frontier models can create pretty good alpha level products. That doesn't mean it is actually an alternative as it is currently a buggy mess and with a lot of purely vibe coded products getting past that initial proof of concept and polishing it up to an actual solid working product is one of those things that remains illusive.
To get ahead of a certain crowd. No, I am not talking about you who actually looks at the output, reviews it and still applies fairly decent basic principles of software development. I am talking about vibe coding in the way it was originally coined, asking the model to do a thing and blindly trusting the output.
Which I have no issue claiming this product is when you see claims like " targets 100% parity in one month". There simply is no way to do this other than blindly trusting all your LLM tooling to work it out.
Edit:
Looking at the commit history shows how insane this is. Commits every few minutes simply means there is no human in the loop. It is all based on the hopes and dreams of people yelling into their prompts and telling the agents to please do a good job.
How many of the programs did you try? I tried the illustrator one, and the photoshop one, and got some work done. I didn't notice it being a particularly buggy mess, but this will largely depend on the features you care about I suppose. There was some old quote I forget but like, each person only uses 10% of word, but they all use a different 10% or something like that.
I tried the illustrator one, and the photoshop one, and got some work done.
With the photoshop one I had weird selection issues across layers. There was also certainly something up with a variety of color tools.
It also is pretty much besides the point as far as I am concerned. With the insane commit frequency I am seeing I really do have my doubts how much human interaction is involved. Which, even with how impressive LLM driven development has gotten is still very much essential. LLMs still struggle with overall cohesion, consistency and generally avoid regression. A thing that is fixed by one prompt might be broken by another and approaches to testing and other things will start to drift over time.
I am not saying that LLM build applications can't work. Which a lot of people replying to me are saying. What I am saying is that you need a lot of human involvement still to make it work. And even then I have only really seen it work on smaller green field projects and have yet to see larger projects like this reach anything like a stable release and remain stable.
There was some old quote I forget but like, each person only uses 10% of word, but they all use a different 10% or something like that.
Also this, it makes the claim of reaching 100% feature parity in a month questionable at best.
I tried the Lightroom clone when it came out a week ago and found a bunch of stuff missing. Raised an issue on Github and the next day the feature was implemented. I'm not sure if they really will succeed in perfectly replicating the Adobe tools, but the pace here is astounding. And I feel like I can meaningfully contribute to these projects just by writing detailed issues which get addressed quickly.
More along the lines of Microsoft Word's dominance, or Excel's - it has thousands of features, of which nobody will use every one of. We can't (generally) recommend Libreoffice as a drop-in replacement though, because the 10% of Word's features that I use work, and the 10% of Word's features you use are broken.
A cleaner phrase might be "People only use 10% of Photoshop's feature set. The problem is that everyone uses a different 10%"
It's not just that LLMs can't (currently?) write maintainable code (though that is a huge issue that's gonna be biting more and more people in the ass) - it's that LLMs let you get to the "finished software" part without the accompanying "people who give a shit" part.
I find vibe coding to be sufficient for smaller projects where you can finish the entire project within a few hours or days. But I spend most of my time on QA. Human QA as well as LLM QA. Claude Opus 5.5 is actually pretty good at coming up with UI/UX/code QA issues. Almost all of the issues it finds are worth fixing - but it certainly is not finding all issues.
One task I've done a few times recently is ask Claude to open a web app in a headless browser and identify UI/UX issues. I have it then fix it and take cropped before/after screenshots. Finally I have it compose a self-contained webpage with every issue and the before/after images, problem statement, proposed solution, and code diff along with "Accept"/"Reject" buttons. 9/10 are both correctly identified issues and good solutions with clean code. I've done the same but with micro code refactors where I have it diagram the dependency tree changes.
For the couple of things I've recently vibe coded (although I do look at the code a bit, but less than I did a few months ago) it's like 40% of the time is adding features and 60% is QA.
I feel like this is underselling how easy it becomes for someone to establish a working product that threatens an industry standard. Moreso if you're able to use that springboard to continue development.
It's of no matter that this is vibecoded. Because now the opportunity exists to reach near feature parity almost instantly. What if next time someone with some programming acumen decides to do this? It became a lot easier to become a real rival to industry standard software packages and the financial pay-for-entry that locked up certain areas of software has been erased.
This was unthinkable a year ago. I'd worry about next year.
I feel like this is underselling how easy it becomes for someone to establish a working product that threatens an industry standard.
I don't think it threatens an industry standard. There have been alternatives to Photoshop and a lot of the suite and that didn't change a single thing. Affinity before they changed their licensing comes to mind. It was affordable, genuinely had a claim at offering most what Photoshop did and was largely ignored by the industry.
I'd also like to point out that outside of the industry there already are alternatives that do a lot. And, which to frank are good enough for 99% of personal use cases. For photoshop there are a variety of programs covering a variety of use cases like Gimp (with PhotoGIMP if you really want photoshop), Krita and Photopea as well as a few others I left out. For vector editing there are Inkscape and Graphite. For raw editing there is darktable, RawTherapee and RapidRAW.
Because now the opportunity exists to reach near feature parity almost instantly.
I think I have a different idea of feature parity than you have. For me it includes stability, both in it being bug free but also in the functionality being stable and working largely the same in a next release. If there is one thing with LLM driven development I have seen is that if not properly governed LLMs tend to suck at keeping functionality stable. They suck at cohesive memory for projects and the only way around that is by having a lot of documentation, memories and constant eye for detail from humans in the loop.
Given the insane commit history on the github repo I sincerely doubt the latter being the case here.
What if next time someone with some programming acumen decides to do this?
People with programming acumen have been doing this. They are getting decent results, generally speaking by tackling smaller green field projects or scoped changes at bigger existing projects. Generally speaking moving at a slower deliberate pace, practicing good development practices and making LLMs follow those practices.
it became a lot easier to become a real rival to industry standard software packages and the financial pay-for-entry that locked up certain areas of software has been erased.
Sort of, do we know how many tokens are spend here? What the actual cost of development was, how much of it this person has eaten out of pocket? We are just trading one locked in eco system for another here.
All things considered, it is possible that I am entirely wrong here. That the project moves forward, finds a balanced approach, reaches a stable point with complete feature parity as they claim. From what I have seen, looking at both this project and across the industry including my own experiments, I highly doubt that will be the case.
What I am 100% standing behind, no matter what, is that the vibe amount is high in this project.
What I am 100% standing behind, no matter what, is that the vibe amount is high in this project.
I don't know what you mean. It only committed ~3000 LOC across 16 commits over the past hour. That's less than 1 line / second. Perfectly possible for a solodev (obligatory /s).
They didn't make something new though - they reverse engineered an existed code base. The tools are impressive and interesting, but it's stealing a code base using tools that stole from everyone elses code base. It's stealing all the way down. You can use adobe for free now? Cool, you do that before with pirating. Did something new happen here?
I don't think there's any indication that they reverse engineered anything. Not to mention that reverse engineering is explicitly legal.
You can see traces in the repo of how they iterated and developed the applications (or rather, directed claude to develop the application) and it's pretty clear they were not using any adobe source code, reverse engineered or not, as reference.
Now, whether or not it's useful is another question - there's already plenty of adobe alternatives that do 95% of what adobe products do - it's just that the 5% is what matters (and also, adobe products are not actually that expensive for a professional, so it's not really worth the hassle).
Clean room reverse engineering is legal, yes. It's widely understood that having the same people do the reversal as the implementation from the derived spec is a good way to get your ass sued into the ground. (This goes way back to things like the cloning of the IBM PC: one team with a test bench poking signals into the system and taking readings, and a fully separate team using their written spec to build a compatible BIOS.)
A straight decompile and port into another language would still be infringing (just as publishing a translation of a novel is infringing). Decompilation is one thing that you never want to do while doing legal reverse engineering, because you're exposing yourself to implementational details.
If no decompilation happened, we get into the untested minefield of "is a leak of the source or a decompilation in the training set?" That's a risk. Any indication that code is sufficiently similar will be used as evidence of infringement.
Regardless, an individual probably can't afford to be sued by Adobe in the first place. This is exactly the kind of case I've been waiting for (though my money was on a Microsoft product)...if this were to go to court, it could establish case law that will make LLM use scarier to lawyers than GPL software. (And then they'll find out that they were trained on GPL code and often emit it verbatim...)
Plus the situation with Photoshop has more to do with being the industry standard and therefore necessary to use in the pro field, and having other features of the creative cloud, not necessary being the best for most people, so this is not that much of a game changer yet. Most pros will likely need to use the actual thing anyway for now (!), to ensure maximum compatibility, avoid chasing creative cloud fonts etc., and non-pros or people who don't need to collaborate with others often already do just fine with Affinity or Photopea.
But I'm telling you if we get something like Solidworks or even Fusion360, that's going to be a game changer even in this state.
Also I already use an app that's afaik 100% vibecoded, it's called Tine and it's a Rust rewrite of Logseq, a personal knowledge management app, which achieves feature parity in selected areas and it works great so far. I reported a bug, it was fixed in a day and included in an update 4 days later. Claude communicated with me in github issues on behalf of the owner.
I reported a bug, it was fixed in a day and included in an update 4 days later. Claude communicated with me in github issues on behalf of the owner.
This is simultaneously amazing and terrifying. I love the responsiveness and quick turnaround. But what happens when someone files “bug reports” that guide Claude in completely out-of-scope (or malicious) directions? Does it just do what it’s told or does it actually have some concept of a product vision in mind that it’s trying to hold to? Seems like a good piece of software could go bad very quickly this way.
Does it just do what it’s told or does it actually have some concept of a product vision in mind that it’s trying to hold to?
The dude who's building Tine seems to be very smart and doing things in seemingly quite sophisticated ways, so I don't doubt that there are multiple gateways to prevent some sort of subversion, but I don't know how resilient they are and I'd assume nobody really does for sure.
Anecdotally I've been working on a fully vibecoded videogame for the last few weeks and a longterm goal/vision is a part of multiple docs files that the agents read every time they start working on something, and usually they make the right decisions, or at least decisions that are not wrong based on that. And I am definitely not experienced in working with agents, this is my first non-tiny project.
This project has gotten a lot of attention. I don't know if the developers will succeed in getting it to something like parity, let alone stability.
But eventually someone will. Probably a small team of someones. And when that happens, regardless of what you think about AI, it will be great for creatives.
Adobe was an enshittification pioneer but their products have remained the de-facto standards in photo editing and vector illustration and no alternative has yet managed to be enough like the originals to be a mass market replacement.
And then there are PDFs, a format seemingly designed by demons to test the sanity of the human race. And still, the defacto standard we can't ever seem to leave behind.
Even just being able to reliably open and save Adobe file formats without losing anything would be a huge win for creatives that can't, or don't want to, pay Adobe's absurdly high prices.
I don't love vibecoding, but a viable, open source, Adobe alternative would still be a net benefit for the world.
Note: I know about the existing alternatives, I use some of them. They aren't good enough. GIMP, for example, has stayed niche for decades for good reasons.
I think this is the real takeaway. It’s early yet. Today’s vibe-coded slopware is better than last year’s. Next year’s will be better still. This specific project isn’t going to unseat Adobe, but that doesn’t mean Adobe’s days aren’t numbered. Eventually the ability for anyone to make their own “good enough” software is going to profoundly disrupt the giants’ ability to keep people dependent on their bloated, overpriced subscriptions. There’s change in the air.
I think the bigger play at hand is: corporations are fine with stealing IP and forcing adoption of AI as long as it’s making money for them. Not so funny when it’s all THEIR IP being stolen and remixed.
I guess now we might see some actual discussion about the ethics and legality of AI use. Of course, it will only be to protect business interests, but still…
I can look for a source later, but discussion on twitter about this was that they decompiled the adobe suite using LLMs (with something like ghidra mcp), then rewrote it in rust. So it should be closely approximating the actual adobe implementation (if I’m not making this up 🤞)
I'm pretty skeptical of that. For one, the author never stated as such. Second, taking the decompilation and then rewriting it in Rust doesn't really make any sense.
1.1 Avoiding GIMP's hole
The criticism of GIMP is historically true, and much of it still applies in 2026.
Bit depth.
GIMP's original core hard-coded 8-bit-per-channel tiles.
High bit depth needed a whole new engine, GEGL. GEGL began around 2000, was introduced as optional in 2.6 (2008), and the port only finished in GIMP 2.10 (2018), "the result of six years of work" on that port alone.
Only then did GIMP get 16/32-bit integer and float processing.
Non-destructive editing.
Photoshop has had adjustment layers since about 1994.
GIMP got non-destructive filters in 3.0 (March 2025), about seven years after 2.10 because of the GTK3 port.
It got non-destructive layer types (link layers, vector layers) in 3.2 (March 2026).
CMYK.
GIMP 3.x is still "a 100% RGB editor" internally.
CMYK exists only as export (JPEG/TIFF/PSD) and soft-proofing.
PSD.
Import is still lossy. Photoshop adjustment layers and most layer styles are not carried over faithfully, and text often needs re-typing.
3.2 converts legacy drop and inner shadows into GIMP filters, and added PSB export.
That isn't really isn't something you'd instruct the agent to if it were actually just does a transpilation. It's not like you can accidentally become GIMP doing that. You'd only need this when you're doing it from scratch and really, really don't want it to accidentally become GIMP.
I mostly agree now, nice hunting. And also I could not find a decent source alleging decompilation. Rather, I found lots of people saying “no bro it wasnt decompilation, the github says it was a clean room implementation!”.
To that, I’d say ofc it makes sense for the author to say that. To say otherwise or admit to decompiling would be opening oneself up to litigation (I assume?) from a very litigious company.
That gimp chat snippet is pretty convincing tho so I’ll choose to believe that this is clean room too. Cheers
Just to clarify, the clean room distinction is mainly about whether or not the developer had access to proprietary code obtained through illegal means. It's relevant when, say, you're developing an emulator of a console which had a data-leak that exposed internal details - if none of the developers have ever seen the leak, you're good to go, whereas if you have, it's likely illegal.
Reverse engineering is completely legal. Decompilation is legal. It's clean room as long as you did not have any leaked source code or documentation you're going off of.
Honestly I might actually sit down and vibe code a few titan killers just to shake shit up and get them to take AI regulation seriously. Just call me Zeus.
The latest Claude and GPT models have done some fucking crazy things, like reverse engineering and splicing two games together. AI is actually one of the main contributors on projects like KytyPS5 (a PlayStation 5 emulator that has made a tonne of progress lately) and AnyPS5 (a tool for porting PS5 games to Windows and Linux), and is a core reason why a lot of recent video game decompilations have taken weeks instead of years to produce.
Am I right to be genuinely concerned by this?
I mean even if you have a raging hate-boner for greedy monopolistic companies like Autodesk and Adobe, it should be alarming that AI is currently in the state where it can be used to create working vibe-slop clones of proprietary software.
I once envisioned a "watershed moment" for AI where someone straight-up recreates Adobe Creative Suite and releases a free working clone of it a few days ago. I was not expecting this moment to have already passed.
I would not be surprised in the slightest if Anthropic and OpenAI get hit with massive lawsuits in the coming months. Heck, I can even see this getting laws rewritten and legal precedent on fair-use and reverse engineering to be overturned.
I mean even if you have a raging hate-boner for greedy monopolistic companies like Autodesk and Adobe, it should be alarming that AI is currently in the state where it can be used to create working vibe-slop clones of proprietary software.
Of course it should be, because we are replacing one type of company for another type of company since this is done with proprietary frontier models. And these companies already have shown all the hallmarks of simply not caring about customers other than as an income stream.
I was not expecting this moment to have already passed.
I mean it hasn't, I am impressed by how far it has gotten. But as I wrote elsewhere it doesn't have near feature parity yet, plenty of bugs and all the hallmarks of a vibe project that is unlikely to leave alpha stage ever.
Open weight models have been right behind closed frontier models for a long time now. I would bet Kimi k3 could probably already do most of this. And for what astra and fable still do better at, just wait a month or two for Kimi k3.5 or k4 or a new deepseek model.
Spend plenty of time learning product and design skills. That's always mattered. Try to be in a position where you can straddle the full pipeline from customer interviews and market evaluation to database optimization. If you can't find work anywhere in that range then software (not just software engineering) has become a completely solved problem.
I don't think these type of cowboy projects are sustainable. There is no way the author is going to continuously support the software for other people, let alone defend it from legal trouble. It is also a matter of time before these kind of projects are mostly done by people trying to distribute malware.
The industry is changing. These things don’t happen all at once or in every company according to a common timeline. Some people will feel the effects before others do, and that’s largely a matter of luck.
My SE career of 18 years ended last year when I and many in my department were unceremoniously laid off. I thought my department head was being unusually gruff when he gave me the bad news and bolted, only later did I learn that he’d been given the axe himself moments before, and his final duty with the company was to do the same to his subordinates. A shitty day all around. AI was never the stated reason, just nebulous “restructuring.” Okay. We were early adopters of Claude Code, so it was at least a factor, but saying more than that is idle speculation.
Anyway, I had already been grappling with career burnout for a few years. The work just wasn’t fun like it had been a decade before. It paid well, but (particularly after the pandemic killed off IRL coworker friendships and cool office perks) it was losing its luster. Then the introduction of Claude workflows redefined my role from a solution finder to a prompt-writer and AI babysitter. I was growing steadily less satisfied with that lot by the time the layoff came, and that was the final straw for me. Of course, I still had to look for new work.
When I re-entered the tech job market I noticed a lot of things were different from the last time I had been there:
Most of the roles now had “Agentic” or “AI-native” in the title. These were prompt-writing and AI babysitting jobs. Not what I ever signed up for or was good at.
Constant ghosting is apparently the norm now. Supposedly that’s because companies are overrun with LLM-slop applications and can’t sort the signal from the noise; also I guess a lot of job boards are now filled with fake ads and scams so the problem is multiplied. Very seldom did I hear back from any listing I applied to.
Pay on offer was notably reduced, across the board, from what I’d been making for years.
One of the things I’ve come to dread the most in job hunting life is the tech interview loop. I’ve been through too many to count, over the years, and have some real trauma from them. I’m not sure how the AI era has changed them, because I didn’t make it that far with anyone this time around. Felt like a slap in the face after all the years of hard work I’ve done.
Anyway to make a long story short, I pivoted. Left the tech sector entirely, and I’m happier for it. I got an administrative job at a large public university. Now I’m doing work that directly helps students. I meet with them and solve problems for them and I get to see the gratitude on their faces. I never got that from any of my old jobs, where I was kept as far away from humans as possible and the ultimate goal of every task I completed was just to make arrow go up.
I’m making about 1/4 of what I used to make, which legitimately sucks, but I’m happy with what I’m doing. Compared to constantly reviewing Claude’s verbose PRs, which were always so confident yet wrong in subtle but critical ways… this new job is mentally so much easier. And my non-technical coworkers think I’m a genius when I show them how to update their email signatures, lol. I feel appreciated, and for once like I might have lasting job security too.
A nice bonus is that I do have access to basically unlimited AI resources, courtesy of the university. So I’m able to keep up with the tech (which is still interesting to me) at no personal cost, and research uses for it on my own terms, to build things that help people, with no PMs breathing down my neck or KPIs driving everything I do. I’m pleased with how things worked out.
But, to summarize — I absolutely do credit AI’s impact on the tech industry as the killing blow to my SE career. I think maybe it hit me earlier than others (including those it’s still coming for but hasn’t reached yet), just because of the nature of my preexisting burnout and middle age long career history. I think many in the field will remain employed, but have to accept that the work they’re doing is irreversibly changing. Maybe they’ll like it better this way, I didn’t. I don’t know what the future holds for me or my SE compatriots, but I can say right now that I’m glad I got out.
I'm currently unemployed (self-employed? I am starting a new business) and it's interesting hearing the difference in opinions on modern software engineering coming from people recently employed in software. From my perspective AI software development is an incredible improvement to the most important reason to create software - finding and solving problems in a cost-effective manner.
I suspect that the pre-existing companies, teams, and products aren't well suited to the new world we've been in for about a year. A million managers, CEOs, or other holders of power are standing on quicksand. And in a panic they demand more from their subordinates, lay people off, and make bad product decisions. But I don't see why, from a fresh outside perspective, I should be upset with LLMs in software engineering. I can create more solutions to more problems in less time and to a higher degree of quality. I have built tools in a week, working less than full time, that would have previously taken a month or more. I can work in domains I previously would have needed months of study to operate within.
I suspect that once the dust settles and the current sand castles collapse and need to be rebuilt there will be plenty of jobs for those able and willing to operate in the new natural language age. People pay for solutions to their problems. And we're more equipped to solve problems than before. I can understand that doesn't solve any short-term issues when someone loses their job during the "dust settling" and "collapsing" phase.
There's no guarantee the jobs will come back. I think that's what a lot of people are concerned about. If a user can describe the problem they're having, and the LLM can just craft them a solution based on that, you don't necessarily need a person to solve the problem.
At some level of AI skill there will be no jobs left that could have been done purely by a human using a computer. My accelerationist side actually looks forward to that. I expect the change to be so intense and so fast that people will finally demand more from their governments.
I don't think that the devaluing of labor has ever, even once in all of history, ended well for those whose labor was devalued. I don't see how the correct prior here is anything other than "this will not end well".
I think of people as highly adaptable. I think there will be a great opening of opportunities with the changes. First mover advantage is real and software engineers still have all sorts of mental skills that are useful. There will be new classes of problems following from the old ones being solved, and I suspect they will be adjacent and thus our skills will be transferable.
So far job prospects still seem fine as these AI tools create so much work and turbulence in projects that need to be managed that I am still as busy as ever, but I find it hard to predict where we will be even 2 years from now let alone for someone just entering the field hoping to get a full career out of it.
creesch | a day ago
Let me fix that title
I am not going to deny it is pretty impressive that current frontier models can create pretty good alpha level products. That doesn't mean it is actually an alternative as it is currently a buggy mess and with a lot of purely vibe coded products getting past that initial proof of concept and polishing it up to an actual solid working product is one of those things that remains illusive.
To get ahead of a certain crowd. No, I am not talking about you who actually looks at the output, reviews it and still applies fairly decent basic principles of software development. I am talking about vibe coding in the way it was originally coined, asking the model to do a thing and blindly trusting the output.
Which I have no issue claiming this product is when you see claims like " targets 100% parity in one month". There simply is no way to do this other than blindly trusting all your LLM tooling to work it out.
Edit:
Looking at the commit history shows how insane this is. Commits every few minutes simply means there is no human in the loop. It is all based on the hopes and dreams of people yelling into their prompts and telling the agents to please do a good job.
capricious_insect | a day ago
How many of the programs did you try? I tried the illustrator one, and the photoshop one, and got some work done. I didn't notice it being a particularly buggy mess, but this will largely depend on the features you care about I suppose. There was some old quote I forget but like, each person only uses 10% of word, but they all use a different 10% or something like that.
creesch | 13 hours ago
With the photoshop one I had weird selection issues across layers. There was also certainly something up with a variety of color tools.
It also is pretty much besides the point as far as I am concerned. With the insane commit frequency I am seeing I really do have my doubts how much human interaction is involved. Which, even with how impressive LLM driven development has gotten is still very much essential. LLMs still struggle with overall cohesion, consistency and generally avoid regression. A thing that is fixed by one prompt might be broken by another and approaches to testing and other things will start to drift over time.
I am not saying that LLM build applications can't work. Which a lot of people replying to me are saying. What I am saying is that you need a lot of human involvement still to make it work. And even then I have only really seen it work on smaller green field projects and have yet to see larger projects like this reach anything like a stable release and remain stable.
Also this, it makes the claim of reaching 100% feature parity in a month questionable at best.
Octofox | an hour ago
I tried the Lightroom clone when it came out a week ago and found a bunch of stuff missing. Raised an issue on Github and the next day the feature was implemented. I'm not sure if they really will succeed in perfectly replicating the Adobe tools, but the pace here is astounding. And I feel like I can meaningfully contribute to these projects just by writing detailed issues which get addressed quickly.
Blakdragon | 18 hours ago
Are you thinking of the 90/10 rule? That the first 90% of a project takes up 10% of the time, and the last 10% takes up 90% of the time?
xethos | 18 hours ago
More along the lines of Microsoft Word's dominance, or Excel's - it has thousands of features, of which nobody will use every one of. We can't (generally) recommend Libreoffice as a drop-in replacement though, because the 10% of Word's features that I use work, and the 10% of Word's features you use are broken.
A cleaner phrase might be "People only use 10% of Photoshop's feature set. The problem is that everyone uses a different 10%"
Weldawadyathink | 9 hours ago
My favorite version is this: the first 90% of the project takes 90% of the time. The last 10% takes the other 90% of the time.
creesch | 6 hours ago
You are thinking of the Pareto principle which is usually said to be 80% 20%.
Gourd | 3 hours ago
The ninety-ninety rule, which is more of a joke but often feels accurate :(
capricious_insect | 15 hours ago
Haha no. But that is also a good one.
More like this: https://www.litmos.com/blog/articles/using-all-the-features-beating-the-80-20-rule-with-customer-training/
ogre | a day ago
Long term maintenance is another elusive skill for LLMs. This, like most vibecoded projects, will probably be abandonware within a year.
JRandomHacker | 20 hours ago
It's not just that LLMs can't (currently?) write maintainable code (though that is a huge issue that's gonna be biting more and more people in the ass) - it's that LLMs let you get to the "finished software" part without the accompanying "people who give a shit" part.
teaearlgraycold | 22 hours ago
I find vibe coding to be sufficient for smaller projects where you can finish the entire project within a few hours or days. But I spend most of my time on QA. Human QA as well as LLM QA. Claude Opus 5.5 is actually pretty good at coming up with UI/UX/code QA issues. Almost all of the issues it finds are worth fixing - but it certainly is not finding all issues.
One task I've done a few times recently is ask Claude to open a web app in a headless browser and identify UI/UX issues. I have it then fix it and take cropped before/after screenshots. Finally I have it compose a self-contained webpage with every issue and the before/after images, problem statement, proposed solution, and code diff along with "Accept"/"Reject" buttons. 9/10 are both correctly identified issues and good solutions with clean code. I've done the same but with micro code refactors where I have it diagram the dependency tree changes.
For the couple of things I've recently vibe coded (although I do look at the code a bit, but less than I did a few months ago) it's like 40% of the time is adding features and 60% is QA.
CptBluebear | 22 hours ago
I feel like this is underselling how easy it becomes for someone to establish a working product that threatens an industry standard. Moreso if you're able to use that springboard to continue development.
It's of no matter that this is vibecoded. Because now the opportunity exists to reach near feature parity almost instantly. What if next time someone with some programming acumen decides to do this? It became a lot easier to become a real rival to industry standard software packages and the financial pay-for-entry that locked up certain areas of software has been erased.
This was unthinkable a year ago. I'd worry about next year.
creesch | 13 hours ago
I don't think it threatens an industry standard. There have been alternatives to Photoshop and a lot of the suite and that didn't change a single thing. Affinity before they changed their licensing comes to mind. It was affordable, genuinely had a claim at offering most what Photoshop did and was largely ignored by the industry.
I'd also like to point out that outside of the industry there already are alternatives that do a lot. And, which to frank are good enough for 99% of personal use cases. For photoshop there are a variety of programs covering a variety of use cases like Gimp (with PhotoGIMP if you really want photoshop), Krita and Photopea as well as a few others I left out. For vector editing there are Inkscape and Graphite. For raw editing there is darktable, RawTherapee and RapidRAW.
I think I have a different idea of feature parity than you have. For me it includes stability, both in it being bug free but also in the functionality being stable and working largely the same in a next release. If there is one thing with LLM driven development I have seen is that if not properly governed LLMs tend to suck at keeping functionality stable. They suck at cohesive memory for projects and the only way around that is by having a lot of documentation, memories and constant eye for detail from humans in the loop.
Given the insane commit history on the github repo I sincerely doubt the latter being the case here.
People with programming acumen have been doing this. They are getting decent results, generally speaking by tackling smaller green field projects or scoped changes at bigger existing projects. Generally speaking moving at a slower deliberate pace, practicing good development practices and making LLMs follow those practices.
Sort of, do we know how many tokens are spend here? What the actual cost of development was, how much of it this person has eaten out of pocket? We are just trading one locked in eco system for another here.
All things considered, it is possible that I am entirely wrong here. That the project moves forward, finds a balanced approach, reaches a stable point with complete feature parity as they claim. From what I have seen, looking at both this project and across the industry including my own experiments, I highly doubt that will be the case.
What I am 100% standing behind, no matter what, is that the vibe amount is high in this project.
Barney | 11 hours ago
I don't know what you mean. It only committed ~3000 LOC across 16 commits over the past hour. That's less than 1 line / second. Perfectly possible for a solodev (obligatory /s).
Blakdragon | 18 hours ago
They didn't make something new though - they reverse engineered an existed code base. The tools are impressive and interesting, but it's stealing a code base using tools that stole from everyone elses code base. It's stealing all the way down. You can use adobe for free now? Cool, you do that before with pirating. Did something new happen here?
stu2b50 | 18 hours ago
I don't think there's any indication that they reverse engineered anything. Not to mention that reverse engineering is explicitly legal.
You can see traces in the repo of how they iterated and developed the applications (or rather, directed claude to develop the application) and it's pretty clear they were not using any adobe source code, reverse engineered or not, as reference.
Now, whether or not it's useful is another question - there's already plenty of adobe alternatives that do 95% of what adobe products do - it's just that the 5% is what matters (and also, adobe products are not actually that expensive for a professional, so it's not really worth the hassle).
redwall_hp | an hour ago
Clean room reverse engineering is legal, yes. It's widely understood that having the same people do the reversal as the implementation from the derived spec is a good way to get your ass sued into the ground. (This goes way back to things like the cloning of the IBM PC: one team with a test bench poking signals into the system and taking readings, and a fully separate team using their written spec to build a compatible BIOS.)
A straight decompile and port into another language would still be infringing (just as publishing a translation of a novel is infringing). Decompilation is one thing that you never want to do while doing legal reverse engineering, because you're exposing yourself to implementational details.
If no decompilation happened, we get into the untested minefield of "is a leak of the source or a decompilation in the training set?" That's a risk. Any indication that code is sufficiently similar will be used as evidence of infringement.
Regardless, an individual probably can't afford to be sued by Adobe in the first place. This is exactly the kind of case I've been waiting for (though my money was on a Microsoft product)...if this were to go to court, it could establish case law that will make LLM use scarier to lawyers than GPL software. (And then they'll find out that they were trained on GPL code and often emit it verbatim...)
V17 | a day ago
Plus the situation with Photoshop has more to do with being the industry standard and therefore necessary to use in the pro field, and having other features of the creative cloud, not necessary being the best for most people, so this is not that much of a game changer yet. Most pros will likely need to use the actual thing anyway for now (!), to ensure maximum compatibility, avoid chasing creative cloud fonts etc., and non-pros or people who don't need to collaborate with others often already do just fine with Affinity or Photopea.
But I'm telling you if we get something like Solidworks or even Fusion360, that's going to be a game changer even in this state.
Also I already use an app that's afaik 100% vibecoded, it's called Tine and it's a Rust rewrite of Logseq, a personal knowledge management app, which achieves feature parity in selected areas and it works great so far. I reported a bug, it was fixed in a day and included in an update 4 days later. Claude communicated with me in github issues on behalf of the owner.
balooga | a day ago
This is simultaneously amazing and terrifying. I love the responsiveness and quick turnaround. But what happens when someone files “bug reports” that guide Claude in completely out-of-scope (or malicious) directions? Does it just do what it’s told or does it actually have some concept of a product vision in mind that it’s trying to hold to? Seems like a good piece of software could go bad very quickly this way.
V17 | 23 hours ago
The dude who's building Tine seems to be very smart and doing things in seemingly quite sophisticated ways, so I don't doubt that there are multiple gateways to prevent some sort of subversion, but I don't know how resilient they are and I'd assume nobody really does for sure.
Anecdotally I've been working on a fully vibecoded videogame for the last few weeks and a longterm goal/vision is a part of multiple docs files that the agents read every time they start working on something, and usually they make the right decisions, or at least decisions that are not wrong based on that. And I am definitely not experienced in working with agents, this is my first non-tiny project.
post_below | a day ago
This project has gotten a lot of attention. I don't know if the developers will succeed in getting it to something like parity, let alone stability.
But eventually someone will. Probably a small team of someones. And when that happens, regardless of what you think about AI, it will be great for creatives.
Adobe was an enshittification pioneer but their products have remained the de-facto standards in photo editing and vector illustration and no alternative has yet managed to be enough like the originals to be a mass market replacement.
And then there are PDFs, a format seemingly designed by demons to test the sanity of the human race. And still, the defacto standard we can't ever seem to leave behind.
Even just being able to reliably open and save Adobe file formats without losing anything would be a huge win for creatives that can't, or don't want to, pay Adobe's absurdly high prices.
I don't love vibecoding, but a viable, open source, Adobe alternative would still be a net benefit for the world.
Note: I know about the existing alternatives, I use some of them. They aren't good enough. GIMP, for example, has stayed niche for decades for good reasons.
balooga | a day ago
I think this is the real takeaway. It’s early yet. Today’s vibe-coded slopware is better than last year’s. Next year’s will be better still. This specific project isn’t going to unseat Adobe, but that doesn’t mean Adobe’s days aren’t numbered. Eventually the ability for anyone to make their own “good enough” software is going to profoundly disrupt the giants’ ability to keep people dependent on their bloated, overpriced subscriptions. There’s change in the air.
DoubleSupercool | 22 hours ago
I think the bigger play at hand is: corporations are fine with stealing IP and forcing adoption of AI as long as it’s making money for them. Not so funny when it’s all THEIR IP being stolen and remixed.
I guess now we might see some actual discussion about the ethics and legality of AI use. Of course, it will only be to protect business interests, but still…
stu2b50 | 21 hours ago
What adobe IP is being stolen? It's closed source to begin with, and the project isn't using any visual or branding IP.
I also don't see any evidence that Adobe is particularly mad.
irlappa | 19 hours ago
I can look for a source later, but discussion on twitter about this was that they decompiled the adobe suite using LLMs (with something like ghidra mcp), then rewrote it in rust. So it should be closely approximating the actual adobe implementation (if I’m not making this up 🤞)
stu2b50 | 19 hours ago
I'm pretty skeptical of that. For one, the author never stated as such. Second, taking the decompilation and then rewriting it in Rust doesn't really make any sense.
Finally, you can just look at the traces left behind in the repo and it does not seem like that's how they worked. For example, in the photoshop like architecture agent markdown file, you see this https://github.com/storytold/photocraft/blob/main/docs/architecture.md
That isn't really isn't something you'd instruct the agent to if it were actually just does a transpilation. It's not like you can accidentally become GIMP doing that. You'd only need this when you're doing it from scratch and really, really don't want it to accidentally become GIMP.
irlappa | 12 hours ago
I mostly agree now, nice hunting. And also I could not find a decent source alleging decompilation. Rather, I found lots of people saying “no bro it wasnt decompilation, the github says it was a clean room implementation!”.
To that, I’d say ofc it makes sense for the author to say that. To say otherwise or admit to decompiling would be opening oneself up to litigation (I assume?) from a very litigious company.
That gimp chat snippet is pretty convincing tho so I’ll choose to believe that this is clean room too. Cheers
stu2b50 | 5 hours ago
Just to clarify, the clean room distinction is mainly about whether or not the developer had access to proprietary code obtained through illegal means. It's relevant when, say, you're developing an emulator of a console which had a data-leak that exposed internal details - if none of the developers have ever seen the leak, you're good to go, whereas if you have, it's likely illegal.
Reverse engineering is completely legal. Decompilation is legal. It's clean room as long as you did not have any leaked source code or documentation you're going off of.
Narry | 21 hours ago
Honestly I might actually sit down and vibe code a few titan killers just to shake shit up and get them to take AI regulation seriously. Just call me Zeus.
Bullmaestro | 8 hours ago
The latest Claude and GPT models have done some fucking crazy things, like reverse engineering and splicing two games together. AI is actually one of the main contributors on projects like KytyPS5 (a PlayStation 5 emulator that has made a tonne of progress lately) and AnyPS5 (a tool for porting PS5 games to Windows and Linux), and is a core reason why a lot of recent video game decompilations have taken weeks instead of years to produce.
Am I right to be genuinely concerned by this?
I mean even if you have a raging hate-boner for greedy monopolistic companies like Autodesk and Adobe, it should be alarming that AI is currently in the state where it can be used to create working vibe-slop clones of proprietary software.
I once envisioned a "watershed moment" for AI where someone straight-up recreates Adobe Creative Suite and releases a free working clone of it a few days ago. I was not expecting this moment to have already passed.
I would not be surprised in the slightest if Anthropic and OpenAI get hit with massive lawsuits in the coming months. Heck, I can even see this getting laws rewritten and legal precedent on fair-use and reverse engineering to be overturned.
creesch | 6 hours ago
Of course it should be, because we are replacing one type of company for another type of company since this is done with proprietary frontier models. And these companies already have shown all the hallmarks of simply not caring about customers other than as an income stream.
I mean it hasn't, I am impressed by how far it has gotten. But as I wrote elsewhere it doesn't have near feature parity yet, plenty of bugs and all the hallmarks of a vibe project that is unlikely to leave alpha stage ever.
Weldawadyathink | 5 hours ago
Open weight models have been right behind closed frontier models for a long time now. I would bet Kimi k3 could probably already do most of this. And for what astra and fable still do better at, just wait a month or two for Kimi k3.5 or k4 or a new deepseek model.
Aldehyde | 22 hours ago
As a CS new grad, this is depressing. Currently in a pretty good position, but at this rate who knows how long that’ll last.
teaearlgraycold | 21 hours ago
Spend plenty of time learning product and design skills. That's always mattered. Try to be in a position where you can straddle the full pipeline from customer interviews and market evaluation to database optimization. If you can't find work anywhere in that range then software (not just software engineering) has become a completely solved problem.
parsley | 12 hours ago
I don't think these type of cowboy projects are sustainable. There is no way the author is going to continuously support the software for other people, let alone defend it from legal trouble. It is also a matter of time before these kind of projects are mostly done by people trying to distribute malware.
[OP] cloud_loud | 22 hours ago
Eh, I've been hearing stuff about how going into computer science means low job prospects for the past like five years now.
balooga | 20 hours ago
I mean, does that make it not true?
The industry is changing. These things don’t happen all at once or in every company according to a common timeline. Some people will feel the effects before others do, and that’s largely a matter of luck.
My SE career of 18 years ended last year when I and many in my department were unceremoniously laid off. I thought my department head was being unusually gruff when he gave me the bad news and bolted, only later did I learn that he’d been given the axe himself moments before, and his final duty with the company was to do the same to his subordinates. A shitty day all around. AI was never the stated reason, just nebulous “restructuring.” Okay. We were early adopters of Claude Code, so it was at least a factor, but saying more than that is idle speculation.
Anyway, I had already been grappling with career burnout for a few years. The work just wasn’t fun like it had been a decade before. It paid well, but (particularly after the pandemic killed off IRL coworker friendships and cool office perks) it was losing its luster. Then the introduction of Claude workflows redefined my role from a solution finder to a prompt-writer and AI babysitter. I was growing steadily less satisfied with that lot by the time the layoff came, and that was the final straw for me. Of course, I still had to look for new work.
When I re-entered the tech job market I noticed a lot of things were different from the last time I had been there:
job huntinglife is the tech interview loop. I’ve been through too many to count, over the years, and have some real trauma from them. I’m not sure how the AI era has changed them, because I didn’t make it that far with anyone this time around. Felt like a slap in the face after all the years of hard work I’ve done.Anyway to make a long story short, I pivoted. Left the tech sector entirely, and I’m happier for it. I got an administrative job at a large public university. Now I’m doing work that directly helps students. I meet with them and solve problems for them and I get to see the gratitude on their faces. I never got that from any of my old jobs, where I was kept as far away from humans as possible and the ultimate goal of every task I completed was just to make arrow go up.
I’m making about 1/4 of what I used to make, which legitimately sucks, but I’m happy with what I’m doing. Compared to constantly reviewing Claude’s verbose PRs, which were always so confident yet wrong in subtle but critical ways… this new job is mentally so much easier. And my non-technical coworkers think I’m a genius when I show them how to update their email signatures, lol. I feel appreciated, and for once like I might have lasting job security too.
A nice bonus is that I do have access to basically unlimited AI resources, courtesy of the university. So I’m able to keep up with the tech (which is still interesting to me) at no personal cost, and research uses for it on my own terms, to build things that help people, with no PMs breathing down my neck or KPIs driving everything I do. I’m pleased with how things worked out.
But, to summarize — I absolutely do credit AI’s impact on the tech industry as the killing blow to my SE career. I think maybe it hit me earlier than others (including those it’s still coming for but hasn’t reached yet), just because of the nature of my preexisting burnout and
middle agelong career history. I think many in the field will remain employed, but have to accept that the work they’re doing is irreversibly changing. Maybe they’ll like it better this way, I didn’t. I don’t know what the future holds for me or my SE compatriots, but I can say right now that I’m glad I got out.teaearlgraycold | 19 hours ago
I'm currently unemployed (self-employed? I am starting a new business) and it's interesting hearing the difference in opinions on modern software engineering coming from people recently employed in software. From my perspective AI software development is an incredible improvement to the most important reason to create software - finding and solving problems in a cost-effective manner.
I suspect that the pre-existing companies, teams, and products aren't well suited to the new world we've been in for about a year. A million managers, CEOs, or other holders of power are standing on quicksand. And in a panic they demand more from their subordinates, lay people off, and make bad product decisions. But I don't see why, from a fresh outside perspective, I should be upset with LLMs in software engineering. I can create more solutions to more problems in less time and to a higher degree of quality. I have built tools in a week, working less than full time, that would have previously taken a month or more. I can work in domains I previously would have needed months of study to operate within.
Knockout_Mouse | 18 hours ago
I don't think people are upset with the LLMs in software. I think they're upset they won't be able to pay the rent or provide for their families.
teaearlgraycold | 18 hours ago
I suspect that once the dust settles and the current sand castles collapse and need to be rebuilt there will be plenty of jobs for those able and willing to operate in the new natural language age. People pay for solutions to their problems. And we're more equipped to solve problems than before. I can understand that doesn't solve any short-term issues when someone loses their job during the "dust settling" and "collapsing" phase.
Knockout_Mouse | 18 hours ago
There's no guarantee the jobs will come back. I think that's what a lot of people are concerned about. If a user can describe the problem they're having, and the LLM can just craft them a solution based on that, you don't necessarily need a person to solve the problem.
teaearlgraycold | 18 hours ago
I think that depends on the class of problem.
Knockout_Mouse | 18 hours ago
And the proliferation and skill of the AI.
teaearlgraycold | 18 hours ago
At some level of AI skill there will be no jobs left that could have been done purely by a human using a computer. My accelerationist side actually looks forward to that. I expect the change to be so intense and so fast that people will finally demand more from their governments.
heraplem | 14 hours ago
I don't think that the devaluing of labor has ever, even once in all of history, ended well for those whose labor was devalued. I don't see how the correct prior here is anything other than "this will not end well".
teaearlgraycold | 13 hours ago
I think of people as highly adaptable. I think there will be a great opening of opportunities with the changes. First mover advantage is real and software engineers still have all sorts of mental skills that are useful. There will be new classes of problems following from the old ones being solved, and I suspect they will be adjacent and thus our skills will be transferable.
[OP] cloud_loud | 18 hours ago
I guess this all makes me a bit less self conscious about being a lowly government employee.
Octofox | 16 hours ago
So far job prospects still seem fine as these AI tools create so much work and turbulence in projects that need to be managed that I am still as busy as ever, but I find it hard to predict where we will be even 2 years from now let alone for someone just entering the field hoping to get a full career out of it.