At some point you hit a project size that is too large or has too many interdependencies, and you have to be very careful about how you manage the context and should expect the llm to start generating too much code or subtle bugs.
Once you hit that size, in my opinion, it's usually best to drop back to brainstorming mode, only use the llm to help you with the design, and either write the code yourself, or write the skeleton of the code yourself and have the llm fill it in.
With too much code, llms just don't seem able yet to only add a few more lines of code, make use of existing code, or be clever and replace a few lines of code with a few more lines of code. They nearly always will add a bunch of new abstractions.
But I have no issues with using Claude Code to write code in larger projects, including adapting to existing patterns, it’s just not vibe coding - I architect the modules, and I know more or less exactly what I want the end result to be. I review all code in detail to make sure it’s precisely what I want. You just have to write good instructions and manage the context well (give it sample code to reference, have agent.md files for guidance, etc.)
This is key.
And this is also why AI doesn't work that well for me. I don't know yet how I want it to work. Part of the work I do is discovering this, so it can be defined.
1. Have the ai come up with an implementation plan based on my requirements
2. Iterate on the implementation plan / tweak as needed, and write it to a markdown file
3. Have it implement the above plan based on the markdown file.
On projects where we split up the task into well defined, smaller tickets, this works pretty well. For larger stuff that is less well defined, I do feel like it's less efficient, but to be fair, I am also less efficient when building this stuff myself. For both humans and robots, smaller, well defined tickets are better for both development and code review.
Once your codebase exceeds a certain size, it becomes counter-productive to have code that is dependent on the implementation of other modules (tight coupling). In Claude Code terms this means your current architecture is forcing the model to read too many lines of code into its context which is degrading performance.
The solution is the same as it is for humans:
"Program to an interface, not an implementation." --Design Patterns: Elements of Reusable Object-Oriented Software (1994)
Suddenly, you've freed up large chunks of context and Claude is now able to continue making progress.
Of course, the project could continue to grow and the relatively small interface declarations could become too many to fit in context. At that point it would be worthwhile taking a look at the application to see if larger chunks of it could be separated from the rest. Managing the number and breadth of changes that Claude is tasked with making would also help since it's unlikely that every job requires touching dozens of different parts of the application so project management skills can get you even further.
The capabilities now are strong enough to mix and match almost fully in the co-pilot range on substantial projects and repos.
These are the perfect size projects vibe coding is currently good for.
So, yes, ONE DAY, AI will be doing all sorts of things (from POTUS and CEO on down), once it is capable of on-the-job learning and picking up new skills, and everything else that isn't just language model + agent + RAG. It the meantime, the core competence of an LLM is blinkers-on (context-on) executing - coding - according to tasks (part of some plan) assigned to it by a human who, just like a lead assigning tasks to human team members, is aware of what it can and can not do, and is capable of overseeing the project.
ARC AGI 2: https://x.com/poetiq_ai/status/2003546910427361402
METR: https://metr.org/blog/2025-03-19-measuring-ai-ability-to-com...
The knowledge necessary to do real work with these things is still mostly locked up in the humans that have traditionally done that work.
You can be explicit about these things.
Why? People don't ask hammers to do much more than bash in nails into walls.
AI coding tools can be incredibly powerful -- but shouldn't that power be focused on what the tool is actually good at?
There are many, many times that AI coding tools can and should be used to create a "small program that already exists in multiple forms in the training data."
I do things like this very regularly for my small business. It's allowed me to do things that I simply would not have been able to do previously.
People keep asking AI coding tools to be something other than what they currently are. Sure, that would be cool. But they absolutely have increased my productivity 10x for exactly the type of work they're good at assisting with.
“It resembles a normal hammer but is outfitted with an little motor and an flexible head part which moves back and forth in a hammering motion, sparing the user from moving his or her own hand to hammer something by their own force and by so making their job easier”
With this context the example may support the quote, nail guns do make driving nails much faster and easier but that's all they do. You can't pull a nail with a nail gun and you can't use it for any of the other things that a regular hammer can do. They do 10x your ability to drive nails though.
On the other hand, LLMs are significantly more multi-purpose than a nail gun.
I think it's for a very reasonable reason: the AI coding tool salespeople are often selling the tools as something other than what they currently are.
I think you're right, that if you calibrate your expectations to what the tools are capable of, there's definitely. It would be nice if the marketing around AI also did the same thing.
And if this submission was an AI salesperson trying to sell something, the comment/concern would be pertinent. It is otherwise irrelevant here.
Because I keep wondering myself if AI is here and our output is charged up, then why am I keep seeing more of the same products but with an "AI" sticker slapped on top of them? From a group of technologists like HN and the startup world, that live on the edge of evolution and revolution, maybe my expectations were a bit too high.
All I see is the equivalent of a "look how fast my new car made me go to the super market, when I'm not too demanding on the super market I want to end up with, and all I want is milk and eggs". Which is 100% fine, but at the end of the day I eat the same omelette as always. In this metaphor, I don't feel the slightest behind, or have any sense of FOMO if I cook my omelette slowly. I guess I have more time for my kids if I see the culinary arts as just a job. And it's not like restaurants suddenly get all their tables booked faster just because everyone cooks omelettes faster.
>It's allowed me to do things that I simply would not have been able to do previously.
You're not the one doing them. Me barking orders to John Carmack himself doesn't make me a Quake co-creator, and even if I micromanage his output like the world's most toxic micromanager who knows better I'm still not Carmack.
On top of that, you would have been able to do previously, if you cared enough to upskill to the point where token feeding isn't needed for you to feel productive. Tons of programmers broke barriers, and solved problems that haven't been solved by anyone in their companies before.
I don't see why everyone claiming that they previously couldn't do something is a bragging point. The LLM's that you're using were trained by the Google results you could've gotten if you Google searched.
Let's not disregard interesting achievements because they are not something else.
Maybe this is possible. Maybe not.
However, it's a fantasy. Granted, it is a compelling fantasy. But its not one based on reality.
A good example:
"AI will probably be smarter than any single human next year. By 2029, AI is probably smarter than all humans combined.” -- Elon Musk
This is, of course, ridiculous. But, why should we let reality get in the way of a good fantasy?
Arguably that's already so. There's no clear single dimension for "smart"; even within exact sciences, I wouldn't know how to judge e.g. "Who was smarter, Einstein or Von Neumann?". But for any particular "smarts competition", especially if it's time limited, I'd expect Claude 4.5 Opus and Gemini 3 Pro to get higher scores than any single human.
No one is propping up a multi-billion dollar tech bubble by promising hammers that do more than bash nails. As a point of comparison that makes no sense.
This is still pretty great!
- https://highload.fun/tasks/15/leaderboard
- https://highload.fun/tasks/24/leaderboard
In both cases my score showed other players that there were better solutions and pushed them to improve their scores as well.
More generally, I do think LLMs grant 10x+ performance for most common work: most of what people do manually is in the training data (which is why there's so much of it in the first place.) 10x+ in those domains can in theory free up more brain space to solve the problems you're talking about.
My advice to you is to tone down the cynicism, and see how it could help you. I'll admit, AI makes me incredibly anxious about my future, but it's still fun to use.
90-99% of an engineer's work isn't entirely novel coding that has never existed before, so by succeeding at what "already exists", it can take us to 10x-100x productivity.
The automation of all that work is groundbreaking in and of itself.
I think that, for a while into the future at least, humans will be relegated to generating that groundbreaking work, and the AI will increasingly handle the rest.
Don't get me wrong, AI is at least as game-changing for programming as StackOverflow and Google were back in the day. Being able to not only look up but automatically integrate things into your codebase that already exist in some form in the training data is incredibly useful. I use it every day, and it's saved me hours of work for certain specific tasks [0]. For tasks like that, it is indeed a 10x productivity multiplier. But since these tasks only comprise a small fraction of the full software development process, the rest of which cannot be so easily automated, AI is not the overall 10x force multiplier that some claim.
That's obviously not going to happen, because AI tools can't solve for taste. Just because a developer can churn out working code with an LLM doesn't mean they have the skills to figure out what the right working code to contribute to a project is, and how to do so in a way that makes the maintainers lives easier and not harder.
That skill will remain rare.
(Also SQLite famously refuses to accept external contributions, but that's a different issue.)
> Contributed Code
> In order to keep SQLite completely free and unencumbered by copyright, the project does not accept patches. If you would like to suggest a change and you include a patch as a proof-of-concept, that would be great. However, please do not be offended if we rewrite your patch from scratch.
I realize that the section, "Open-Source, not Open-Contribution" says that the project accepts contributions, but I'm having trouble understanding how that section and the "Contributed Code" section can both be accurate. Is there a distinction between accepting a "patch" vs. accepting a "contribution?"
If you're planning to update this page to reduce confusion of the contribution policy, I humbly suggest a rewrite of this sentence to eliminate the single and double negatives, which make it harder to understand:
> In order to keep SQLite in the public domain and ensure that the code does not become contaminated with proprietary or licensed content, the project does not accept patches from people who have not submitted an affidavit dedicating their contribution into the public domain.
Could be rewritten as:
> In order to keep SQLite in the public domain and prevent contamination of the code from proprietary or licensed content, the project only accepts patches from people who have submitted an affidavit dedicating their contribution into the public domain.
Found that paperwork here: https://www.sqlite.org/copyright-release.html
I will make sure not to spread that misinformation further in the future!
Update: I had a look in fossil and counted 38 contributors:
brew install fossil
fossil clone https://www.sqlite.org/src sqlite.fossil
fossil sql -R sqlite.fossil "
SELECT user, COUNT(*) as commits
FROM event WHERE type='ci'
GROUP BY user ORDER BY commits DESC
"
Until it decides to include code it gathered from a stackoverflow post 15 years ago probably introducing security related issues or makes up libraries on the go or even worse, tries to make u install libs that were part of a data poisoning attack.
As someone who frequently uses Claude Code, I cannot say that a year's worth of features/improvements have been added in the last month. It bears repeating: if AI is truly a 10x force multiplier, you should expect to see a ~year's worth of progress in a month.
They are by definition a biased source and should not be referenced as such.
I do however think he is not an actively dishonest source. When he says "In the last thirty days, I landed 259 PRs -- 497 commits, 40k lines added, 38k lines removed. Every single line was written by Claude Code + Opus 4.5." I believe he is telling the truth.
That's what dogfooding your own product looks like!
I am happy as is tbh, not even looking for AGI and all. Just that the LLM be close enough to my thinking scale so that it does not feel "why am I talking with this robot".
Not either of the species of algorithms you've described, but still an advance.
This is going to destroy my home network, since I never moved it off the little Lenovo box sitting in my laundry room beside the Eero waypoint, but I’m out of town for three days, so
Granted, the seed of the idea was someone posting about how they wired pyiodide to Ace in 400 lines of JavaScript, so I can’t truly argue it’s non-trivial.
As a light troll to hackernews, only AI-written contributions are accepted
[Edit: the true inception of this project was my kid learning Python at school and trinket.io inexplicably putting Python 3 but not 2 behind the paywall. Alas, Securely will not let him and his classmates actually access it ]
Here is how it works: I take the latest state of the art model, usually one of the two or three currently being hyped....and ask it to create a short document that teaches Java, Python, or Rust, in 30 to 60 min, complete with code examples. Then I ask the same model to review its own produced artifact, for correctness and best practices.
What happens next is remarkably consistent. The model produces a glowing review, confidently declaring the document “production ready”… while the code either does not compile, contains obvious bugs, or relies on outright bad practices.
When I point this out, the model apologizes profusely and generates a “fixed” version which still contains errors. I rinse and repeat until I give up.
This is still true today, including with models like Opus 4.5 and ChatGPT 5.2. So whenever I read comments about these models being historic breakthroughs, I can’t help but imagine they are mostly coming from teams proudly generating technical debt at 100× the usual speed.
Things go even worst, when you ask the model to review a Cloud Architecture....
I think if you want to make this a useful experiment you should use one of the coding assistants that can test and iterate on its code, not some chatbot which is optimized to impress nontechnical people while being as cheap as possible to run.
Is that how we call Opus 4.5 now? :-)
"I paid for the $100 Claude Max plan so you don't have to - an honest review" -https://www.reddit.com/r/ClaudeAI/comments/1l5h2ds/i_paid_fo...
I'm wondering if your test includes allowing the models to run their code in order to validate it and then fix it using the error output? Would you be willing to share the prompts and maybe some examples of the errors?
I haven't had many problems working in Claude Code even with full on "vibe coding". One notable recent exception was in writing integration tests for a p2p app that uses WebRTC, XTerm.js, and Yjs where it ran into some difficulty creating a testing framework that involved a headless client and a local MQTT broker where we had to fork a few child processes to test communication between them. Opus got bogged down working on its own so I stepped in and got things set up correctly (while chatting with Opus through the web interface instead of CC). The problem seemed to be due to overfilling the context since the test suite files were too long so I could have probably avoided the manual work by just having Opus break those up first.
That seems to be a strawman here, no? Sure, there exist people/companies claiming 10x-100x productivity improvements. I agree it's bullshit.
But the article doesn't seem to be claiming anything like this - it's showing the use of vibe-coding for a small personalized side-project, something that's completely valid, sensible, and a perfect use-case for vibe-coding.
1. Current LLMs do much better than produce "small programs that already exist in multiple forms in the training data". Of course the knowledge they use does need to exist somewhere in training data, but they operate at a higher level of abstraction than simply spitting out programs they've already seen whole cloth. Way higher.
2. Inventing a new compression algorithm is beyond the expectations of all but the the most wild-eyed LLM proponents, today.
This is the "promise" that was being sold here and in reality, we yet haven't seen anything innovative or even a sophisticated original groundbreaking discovery from an LLM with most of the claims being faked or unverified.
Most of the 'vibe-coding' uses here are quite frankly performative or used for someone's blog for 'content'.
> Claude did not invent that idea. It executed it.
> Claude handled implementation. I handled taste.
This style of writing always gets me now :)
^^ These dramatic statements are almost always AI influenced, I seem to always see them in people's emails now as well. "we didnt reinvent the wheel. we are the wheel."
Heh.
Seriously: what tool do you want to use that's immediately available to the absolute lowest common denominator "writers" on the Internet?
"It's not X, it's Y" literally makes my stomach churn from seeing so much of it on LinkedIn.
I usually grimace at "GPT smell". The lines you quoted stood out to me as well, but I interpreted them as "early career blogger smell". It's similar, but it didn't come off as AI. I think because it avoided grandiose words, and because it reused the exact same phrase format (like a human tic) rather than randomly sampling the format category (like an AI). This is what human text looks like when it's coming from someone who is earnestly trying to have a punchy, intentional writing style, but who has not yet developed an editor's eye (or asked others to proofread), which would help smooth out behaviors that seem additive in isolation but amateur in aggregate.
Did others share the impression that it's a human doing the same classic tricks that AI is trained to copy, or does anything in this category immediately betray AI usage?
Many weaker or non-native writers might use AI for that "editor's eye" without realizing that they are being driven to sound identical to every other blog post these days. And while I'm certainly growing tired of constantly reading the same LLM style, it's hard to fault someone for wanting to polish what they publish.
I think it some kind of value - vibe dynamics that play in making the brain conscious about it being written with AI or otherwise.
It was very surprising to me that claude wasn't very good at even the latter. It took several rounds of prompt refinement, including very detailed instructions, to get even 75% non-broken links, from ~30% on first attempt. It really liked to hallucinate goodreads URLs with the book title in them but an invalid ID (the title part of URLs is ignored by goodreads).
The former was less surprising... It attempted a very rough manual strategy for generating superimposed SVG outlines on the books, which often started okay but by the right side of the the image often didn't even intersect the actual spines. I tried to verbally coax it into using a third party segmenter, but with no luck. We eventually found a binary search style iterative cropping strategy that worked well but was taking 6 minutes and nearly a couple dollars per spine so I killed that and just did the SVG outlines in figma myself.
* https://andrewblinn.com scroll down to find the bookcase and drag down on it to zoom and unlock the books
I really enjoy the ability to get started quickly with a known idea like “make a single user letterboxd clone” with a system prompt that explains my preferred tech stack. From there it’s relatively easy to start going in and being the tastemaker for the project.
I think people being able to build their own bespoke apps is a huge super power. Unfortunately I don’t think the tools today do a good job of teaching you how to think if you aren’t already a software engineer. Sonnet rarely grasps for an abstraction.
This is such a key thing I remind myself when I build apps like this for myself. I have a similar app that has a page with 900-odd ratings, and another with 550 owned books. I decided that I won't bother with infinite scroll or complex search and filtering until my browser can no longer handle rendering that data. "Find in page" works well enough for me for now.
My experience with Claude was mostly very good. Certainly the UI is far better than what I'd come up with myself. The backend is close to what I'd write myself. When I'm unhappy I'm able to explain the shortcomings and it's able to mostly fix itself. This sort of small-scale, self-contained project was made possible thanks to Claude.
Other times it just couldn't. The validation for the start and end dates it decided was z.string().or(z.date()).optional().transform((val) => (val ? new Date(val) : undefined)). It looked way too complex. I asked if it could be simplified, Claude said no. I suggested z.date().optional(). Claude patiently explained this was impossible. I tried it anyway, it worked. Claude said "you're absolutely right!". But this behaviour was the exception rather than the rule.
cool to check out your version as well thanks for sharing.
So many systems are fault-tolerant, and it’s great to remember that in a world where LLMs introduce new faults. Kudos to OP for this mindset; more anti-AI posters would benefit from sitting with the idea from time to time.
Side note: I once wrote about recreating Delicious Library: https://dingyu.me/blog/recreating-delicious-library-in-2025
I've actually been working on something similar since I also had this pain (plus I'm too cheap to buy all the books I'm reading)
To solve the issue related to having take the photos myself, I scrape from places like eBay to get the accurate spine images. Obviously this isn't 100% accurate, but like you I also decided 90% accuracy is good enough.
This is my experience with agents, particularly Claude Code. It supplies sufficient activation energy to get me over the hump. It makes each next step easy enough that I take it.
However, to me as a person with an anti-library as well, this kind of defies the purpose of having it in the first place. I can't say I browse my books too often but when I want to find something, I'd rather browse physical things on a shelf rather than some out-of-date UI with fetched thumbnails. Of course the organization happens in physical space too: this is why we have shelves, labels and such.
Obviously no judgement or criticism for the author, just sharing thoughts.
I’ve been vibe-coding a personalized outliner app in Rust based on gpui and CRDTs (loro.dev) over the last couple days - something just for me, and in a big part just to explore the problem space - and so far it’s been very nice and fun.
Especially exploring multiple approaches, because exploring an approach just means leaving the laptop working for an hour without my attendance and then seeing the result.
Often I would have it write up a design doc with todos for a feature I wanted based on its exploration, and then just launch a bash for loop that launches Claude with “work on phase $i” (with some extra boilerplate instructions), which would have it occupied for a while.
Something you don’t really mention in the post is why do this? Do you have an end goal or utility in mind for the book shelf? Is it literally just to track ownership? What do you do with that information?
I want my website to slowly become a collection of things I do and like, and this bookshelf is just one of those pieces.
I wonder if you could develop this as an add on to Hardcover.app - you could fetch people's books, images, and display the bookshelf.
All the data seems to be there:
https://hardcover.app/@BenHouston3D/books/read?order=owner_l...
Vibe coded a library last month for my website however its much simpler and has Antilibrary section for all the stuff I have not read.
The whole thing feels a bit like god-of-the-gaps situation, where we keep trying to squeeze humanity into whatever remaining little gaps the current generation of AI hasn’t mastered yet.
you can tell by how many people earnestly share AI generated images, many are completely tasteless but people don't care
I’ve been working on opensource.builders, and what I keep seeing is that proprietary software has incentives that make personal software appealing. Commercial software is built for many different use cases and as it grows, it probably wants to capture more of the market. But with each new use case, new complexity is introduced. At some point, users are “programming,” but it’s just your config. With AI, why not program something that works exactly for your business?
That’s why so many people are getting into vibe coding. They’ve gone through the nightmare of using software where they only need one feature, or tools that have slowly gotten enshittified, and they realize it might actually be easier to build something that fits them instead of wrestling with layers of complexity that exist for everyone else’s edge cases.
Very relatable!
> I started asking for things I did not need.
For a community that prides itself on depth of conversation, ideas, etc. I'm surprised to so much praise for a post like this. I'll be the skeptic. What does it bring to you to vibe code your vibe shelf?
To me, this project perfectly encapsulates the uselessness of AI, small projects like this are good learning or relearning experience and by outsourcing your thinking to AI you deprive yourself of any learning, ownership, or the self fulfillment that comes with it. Unless, of course, you think engaging in "tedious" activities with things you enjoy have zero value, and if getting lost in the weeds isn't the whole point. Perhaps in one of those books you didn't read, you missed a lesson about the journey being more important than the destination, but idk I'm more of a film person.
The only piece of wisdom here is the final sentence:
> Taste still does not [get cheaper].
Though, only in irony.
This is a critical observation of the vibocene.
Vibe coding has really helped me explore skills outside of my comfort zone which can then be applied in combination with other existing skills or interests in new ways.
In the case of your project, I imagine that now that you can gather data such as books from an image of a bookshelf, you can do something similar in infinite other ways.
First, I took photographs of all my physical books simply by photographing the bookshelves such that the book spines were visible.
Then passed the photographs with a prompt akin to, "These are photographs of bookshelves. Create a table of book title and book author based on the spines of the books in these photographed shelves." ChatGPT4’s vision model handled this no problem with pretty high accuracy.
I then vibe-coded a Python program with ChatGPT4 to use the Google Books API (an API key for that is free) to generate a table, and then a CSV, of: book title, book author, and isbn13. Google Books API lets you look up an ISBN based on other metadata like title and author easily.
Finally, I uploaded the enriched CSV into a free account of https://libib.com. This is a free SaaS that creates a digital bookshelf and it can import books en masse if you have their ISBNs. You can see the result of this here for my bookshelf:
https://www.libib.com/u/freenode-fr33n0d3
There are some nice titles in there for HN readers! My admin app for Libib (the one at https://libib.com) is more full-featured than the above public website showcases. It's basically software for running small lending libraries. But, in my case, the “lending library” is just my office’s physical bookshelf.
I also added a Libib collection there that is a sync of my Goodreads history, since I read way more Kindle books than physical books these days. That was a similarly vibe-coded project. But easier since Goodreads can export your book collection, including isbn13, to a file.
As for my actual physical bookshelf, it is more a collection of books I either prefer in print, or that are old, or out-of-print, or pre-digital & never-digitized.
I liked the Libib software so much I end up donating to it every year. I originally discovered it because it is used for Recurse Center’s lending library in the Recurse Center space in Brooklyn, NY (https://recurse.com).
Also, Libib has a Android, iPhoneOS, and iPadOS apps -- these are very basic but they do allow you to add new books simply by scanning their ISBN barcode, which is quite handy when I pick up new items.
I did enjoy reading the OP writeup, it’s a fun idea to vibe-code the actual digital bookshelf app, as well!
I wish book archive sites like archive.org scanned and stored the book spines as well as the covers, but AFAICT none do.
This is the right mindset.
SerpAPI provides a very valuable programmatic access to search that Google are hell bent on never properly providing
One-off scripts and single page html/css/js apps that run locally are fantastically accessible now too.
As someone who doesn't code for a living, but can write code, I would often go on hours/day long side quests writing these kind of apps for work and for my personal life. I know the structure and architecture but lack the fluency for speedy execution since I'm not writing code everyday. Claude code fills that speed gap and turned my days/hours long side quests into minutes for trivial stuff, and hours for genuinely powerful stuff at home and at work.