AI has always been good at recommending things you never thought of, so I get that for sure.
For the other stuff, w/o asking you to do a ton of homework, can you give an example of a feature it wrote from scratch? Was it based on a feature you wrote and then it patterned itself after your coding style and structural choices? Code samples would be awesome. I am really interested to see what it can do.
Also I’d be really curious to see what this prompts file looks like. Is that the only input it needs other than your code?
FWIW your experience is more laudatory than maybe 95% of the reports I’ve read on HackerNews. And a lot of those are from proponents or people who were excited to try the technology. IE - not just skeptics like me.
Sure most of the things I have it do I either have something already there, or something very similar it can pattern itself after, and those typically go very well and it usually gets it right. But it can do large full stack features by itself too as long as you know what you’re doing or how to nudge it.
This prompt is a decent example, my game needs a “draw” or agreed on tied game feature, I set up a few things, had an older dummy file that wasn’t doing anything, and “crafted” this prompt and this was done in 5 minutes.
Then we hacked out the “saves to database” flow and that was fine too. Did it take me a while to write that prompt sure, but this is information that I’d absolutely have to have 100% have internalized anyways to do it myself.
If you mean the AGENTS.md for prompts file yeah I just keep telling it to update it every time it writes code or generally behaves in a way I don’t want it to but those times are rarer and rarer. Here’s my current one. (fuck vercel obv)
At least 10 more examples of features at this scale it was fine at too, and in the past month I’d say. Could I have done the above feature in an hour or two sure, but now I don’t have to. Could a junior dev? No way, a week, probably not at all. It could be that I’m getting massive gains because my project is a one-language full stack monorepo so all of the context is there, but that’s not super unusual. I’m surprised to hear other people not having this type of success tbh.
That’s pretty impressive. If you ever felt so-inclined you could do a writeup on HackerNews and it might blow up. Maybe there are tons of other solo game/similar project devs out there doing the same thing, but I have yet to see a post that lays it out like that.
Is it getting paid to promote Vercel or something?
Is the structure of the .md file totally free-form by you, or is there a recommended template or something?
Do you have it write integration or end to end tests in addition to unit tests?
2-space indentation FTMFW.
I never heard kebabCase before. It’s not camelCase anymore?
Inject most of the Personal Preferences into my veins. You and I have more similar code tastes than most devs I’ve worked with.
Really would have been nice if the third-party devs whose code I’m slogging through now would have followed half of those. Stuff like not organizing your import statements, while claiming to be a senior dev, is a big red flag to me. Actually I know these guys used a lot of AI because features have either a) no comments and debugging output or b) lots of useless comments and verbose console.logs that start with emojis. Gee, I wonder which the human wrote and which Claude wrote.
Their code has a lot of super head-scratcher design decisions. The app has a bunch of complex multi-page forms with endless variations. So they decided the easiest way to handle this is to keep an object in localStorage that represents the state of the form, then pass that to the back end with each page submit and the final actual save submit.
A few problems though:
They’re trusting stuff like the item price coming from the browser in this big object, and not even looking it up in the database. Would you like to buy a $100 ticket for $1? Just send us a different object. I think this is because they started on the admin side form-builder, where we can trust our internal users not to hack us, then just carried the same paradigm over to the user-facing buy flow. (I’m fixing that obv)
I’ve spent countless hours fighting bugs when the big cached object gets out of sync with what’s in the DB. It’s much better imo to always get a fresh copy of the object from the DB after each save. Yes it’s one extra call, but this isn’t a performance-heavy app.
On the admin side, users can go back and edit things after the initial big save (think Wordpress publish). They clearly never really considered that because it breaks a bunch of their paradigms. And before they could fix it all, we ran out of money (thanks Trump) and they were yanked off the project. So now I have to fix all their bad assumptions.
Anyway, not sure how much of that I can pin on AI. Real devs could easily do stuff like that, especially a consultancy that never has to maintain the stuff they build. I just needed to vent a little.
I was banned from HN years ago for incessently needling not-at-all-prepared-to-deal-with-someone-like-me trumpy techbros so no I won’t be doing that. But I am actually planning on making a youtubez where I write a full stack feature myself, and then have different AI agents try to do the same feature type of experiment and compare/contrast, with almost certainly hilarious results, to coincide with my game launch in a few weeks. Forum enemies and lot lechers, then will be your chance! @spidercrab plz be ready - I want you to be prepared, your hands on the buttons, to do absolutely nothing about it when I inevitably get doxxed here for the ~5th time.
Vercel ceo of course but I don’t really think the nextjs lockin is great anyways
md file is completely generated by AI agents, I don’t write any of this stuff, I say “hey this sucks dont do it this way, write that preference to agents md” and off it goes. But this is a good example of places to start, the second anyone starts up one of these things the first prompt should be “analyze this repo and generate an extensive AGENTS.md file that contains code style preferences and any other useful info agentic coding programs should know before writing code”.
I gave up on tabs vs spaces years ago its all the same now. I used be a big 4 tab guy but maybe vsc UX got better? it doesn’t bother me. IDK about kebab case like how many of these are there? At least I use language that doesn’t really use underscores.
If devs you know are using AI a lot but their code is sloppy its 100% on them. Things like import orders and variable definitions I’ve always been super nitty about and probably spent 10s of hours in the past copy and pasting lines to standardize things, but now AI just does it but you have to pay attention, you have to have it write to agents, you have not be lazy, most people are lazy of course. Stuff like storing browser state of forms in localstorage there’s no way AI would blindly suggest that so that’s all on them. Probably..
I know I’m on an island but I can’t stand closing brackets and else on the same line. It looks too bunched up and I can’t instantly see the blocks. Just thought I’d check if we were really code-style BFFs.
Stuff like organizing import statements really matters imo and it’s not being nitty. The time wasted hunting through those looking for what you want is not trivial (applies to lots of other things that seem nitty imo). I go react/node→other third party→our lib JS→our lib styles (images, other style stuff)→feature-level utils, components, then styles→page or component-level helpers then styles.
Let me know when you make the YouTube and I’ll promote it on HN if you want.
LeCun doesn’t mince his words about why he ultimately decided to leave Meta after more than a decade. Staying became politically difficult, he tells me. And while Zuckerberg likes LeCun’s world model research, the crowd who were hired for the company’s new superintelligence push are “completely LLM-pilled”.
You don’t tell a researcher what to do. You certainly don’t tell a researcher like me what to do
This clearly alienated LeCun. “I’m sure there’s a lot of people at Meta, including perhaps Alex, who would like me to not tell the world that LLMs basically are a dead end when it comes to superintelligence,” he says. “But I’m not gonna change my mind because some dude thinks I’m wrong. I’m not wrong. My integrity as a scientist cannot allow me to do this.
Emphasis mine. I don’t know what his definition of superintelligence is. The Soares book has been heavily criticized but their definition is roughly an AI that is better at everything than all of us put together.
Yeah one thing I don’t use at all is “MCP”, the couple time I’ve glanced at it I don’t get it so I’d love to see a workflow that someone uses that gets value from that for just normal dev work. I did ask claude code and codex this now for fun:
Claude think’s I’m pretty ideal guys, thats amazing!
So yeah probably having everything consistent and shared (i.e. exactly what a monorepo is..) is probably why I’ve been able to have it do so much lately, but this app is mostly done and these things tend to snowball imo.
FWIW these are all things that make life much easier for human devs too and extend the lifespan of the app before things eventually go chaotic (assuming the app keeps growing, which if it’s being used, it should keep growing).
So it sounds like getting AI to work well actually enforces tight design up front, which is cool.
All apps eventually turn into a big slog of legacy cruft. You know you’ve reached it when developers start groaning about new feature requests. On a greenfield app, I try to make design decisions that keep the app “fun” to work on as long for as possible.
This often manifests as a battle with management to do things the right way and not rack up tech debt, which of course takes longer than just barreling out features as fast as possible. Solo devs may feel the same pressure internally, or not.
In the corporate world it takes a strong dev who management trusts to push back like this. Except sometimes management does an end run around by picking other devs, or bringing in a consultancy. Then I often have to clean up their mess, like I’m doing right now.
Yeah I put a lot more stock in what Grue is doing that Kyle, because I know Grue is a real developer. But also Kyle’s vibe-coded slop might be good enough for what he’s doing. If he doesn’t need to build on it and just needs to crank out some driveline website with a few back-end features, maybe he’s fine.
For human devs, I’m still not 100% sold on Typescript for a self-contained app. It just seems to turn Javascript into Java, and slow things down. But maybe for AI it’s best because AI gets much quicker feedback when something doesn’t compile.
One of my New Year’s resolutions is to figure this out. Like build an agentic process with MCP even if it doesn’t work well just so I understand it. I don’t write code very often but was a developer for many years and lead teams of developers along with other IT professionals. We are seeing 25-30% productivity gains from the enthusiastic, high skill, early adopters in my team, and I’ve basically been able to get back into coding a little bit to form my own opinion about vibe coding, ai assisted coding, and ai assisted systems development (eg ai design agent, pm agent, testing agents, ai assisted coding, ai written documentation, etc).
We’ve had some good success but it’s worked much better on greenfield than brownfield apps an AI doesn’t do the job for you. It’s a nice co creator think of it like the old pair programming paradigm. You review the AI code and improve. Develop great user stories. Discuss architecture, potential future evolution, and data architecture in plan mode before asking t to build anything. Ask AI to review your code and suggest improvements. You can ask it to find bugs but not patch them if you’d rather do it yourself.
Another thing I’m skeptical about is any measure of developer productivity more precise than sucks/average/rock star, which a good manager should know anyway. You can look at github commits and LoC as a very rough guide. But if the devs ever get wind, they’ll start gaming it, even subconsciously.
How are you measuring these gains and are the devs aware of it?
Yep best way to put is for $20/mo I’ve hired a senior dev who can be leveraged in many ways but a “shadow” that confirms what you’re doing is not stupid, and finds issues that a human would easily miss in code review is worth it by itself, and it can do a lot more than that.
