Imho your post summarizes 90% of the posts I see about AI coding on HN: not understanding the tools, not understanding their strenghts and weaknesses, not being good at prompting or context management yet forming strong(ish) opinions.

If you don't know what they are good at and how to use them of course you may end up with mixed results and yes, you may waste time.

That's a criticism I have also towards AI super enthusiasts (especially vibe coders, albeit you won't find much here), they often confuse the fact that LLMs often one shot 80% of the solutions with the idea that LLMs are 80% there, whereas the Pareto principle well applies to software development where it's the hardest 20% that's gonna prove difficult.

Hey! I would encourage you to try our Claude code instead, which is also part of your subscription. It's a CLI that takes care of many of the issues you encountered, as it works directly on the code files in a directory. No more copy pasting or unscrambling results. Likewise, it can run commands itself to e.g. compile or even test code.

I'm just shy of 30 years experience. I think I've spent more time learning how to use these tools than any other technology I've learned, and I still don't know the best way to use them.

They certainly weren't a time-saver right away but they became one after some time giving them a real shot. I tested their + my limits on small projects, working out how to get them to do the whole project, figuring out when they stop working and why, figuring out which technology they work best with, figuring out the right size problems to give them, figuring out how to recognize if I'm asking them something they can't do well a ask something different instead, guide them into creating code that they can't actually continue to be successful with.

I started last December in Cursor's agentic mode and have been in Claude Code ever since probably March or April. It's definitely been a huge boost all year for side projects - but only in the last couple months have I been having success in a large codebase.

Even with all this experience I don't know that I would really be able to get much value out of the chat interface for them. They need to be proposing changes I can just hit accept or reject on (this is how both Claude Code and Cursor work btw - you don't have to allow it to write to any file you don't want or execute any command you don't want).

The fact that you use Claude, not gpt5 which is light-years ahead for coding, tells me all I need to know.

I hate to play the "you're holding it wrong" card, but when I started, I had more or less the same experience. Eventually you start to learn how to better talk to it in order to get better results.

Something I've found useful with Claude Code is that it works a lot better if I give it many small tasks to perform to eventually get the big thing done, rather than just dumping the big thing in its lap. You can do this interactively (prompt, output, prompt, output, prompt, output...) or by writing a big markdown file with the steps to build it laid out.

Yeah I’m still not more productive. Maybe 10% more. But it alleviates a lot of mental energy, which is very nice at the age of 40.

I didn't see much mention of tab-completions in the survey and comments here. To me that's the ultimate coding AI is doing at my end, even though it seems to pass unnoticed nowadays. It's massive LOC (and comments!), and that's were I find AI immensely productive.

Can you be specific about what you're asking it to write?

I'm pretty confidient that I couldn't get it to implement a element in a web browser. I'm talking about C++ in WebKit or Chromium, not a custom element in HTML/JS. Let's say browsers wanted a new data-table element that natively implemented a scrolling window such that you registered events to supply elements and it asked for only the portion that were visible. I feel like those code bases are too complex for LLMs to add this (could certainly be wrong).

In any case, more concrete examples would help these discussions.

I can give a concrete example: I just asked ChatGPT (should have asked something more integrated into my editor), to give me JavaScript to directly parse meta data out of an MP4 file in JS in the browser. It gave my typescript but told me I should look at mp4box.

I spent 15-20 minutes getting a mp4box example setup (probably should have asked for that too). Only to find that I think mp4box does not get me the data I wanted.

So I went back to ChatGPT and asked for JavaScript because I was hacking in something like jsfiddle. It gave me that and it worked!

I then said I wanted the title and comments meta data as that was missing from what it gave me. That worked too, first try.

I’m not yet up to half (because my corporate code base is a mess that doesn’t lend itself well to AI)

But your approach sounds familiar to me. I find sometimes it may be slower and lower quality to use AI, but it requires less mental bandwidth from me, which is sometimes a worthwhile trade off.

100%. My instructions are often pretty exact "i want a genserver that fetches the data every X sec and updates Y if new data"

I typed less using AI, so it is a net win I guess

I’ve arrived at something similar:

* brainstorm all the ideas, get Claude to write docs + code for all them, and then throw away the code

* ask it to develop architecture and design principles based on the contents of those docs

* get it to write a concise config spec doc that incorporates all the features, respects the architecture and design as appropriate

* iterate over that for a while until I get it into a state I like

* ask it to write an implementation plan for the config spec

* babysit it as I ask it to implement phase by phase of the implementation plan while adhering to the config spec

It’s a bit slower to than what I’d hoped originally, but it’s a lot better in terms of end result and gives me more opportunity to verify tests, tweak implementation, briefly segue or explore enhancements, etc.

> To get that to work half-decent, you have to take on a PM/Tech-lead role, you're no longer a senior engineer.

But you’re saying it can be half-decent?

The problem is that about 75% of HN commenters have their identities tightly wound up in being a (genuflect) senior engineer and putting down PM/tech-lead type roles.

They’ll do anything to avoid losing that identity including writing non-stop about how bad AI code is. There’s an Upton Sinclair quote that fits the situation quite nicely.

I love the split personality vibe here.

When it comes to dealing with shitty platforms AI is really the best thing ever. I have had the misfortune of writing automations for Atlassian with their weird handling of refresh keys and had AI not pointed out that Atlassian had the genius idea of invalidating refresh keys after single use, I would have wasted a lot more of my time. For this sort of manual labout, AI is the best tool there is.

> Meanwhile many of the seniors around me are stuck in their ways, refusing to adopt interactive debuggers to replace their printf() debug habits, let alone AI tooling...

When I was new to the business, I used interactive debugging a lot. The more experienced I got, the less I used it. printf() is surprisingly useful, especially if you upgrade it a little bit to a log-level aware framework. Then you can leave your debugging lines in the code and switch it on or off with loglevel = TRACE or INFO, something like that.

Interactive debuggers and printf() are both completely valid and have separate use-cases with some overlap. If you're trying to use, or trying to get people to use, exclusively one, you've got some things to think about.

That's funny. I remember using interactive debuggers all the time back in the '90s, but it's been a long time since I've bothered. Logging, reading, and thinking is just... easier.

The right tool for the right job. If someone gets the job done with printf() then that would be good enough for me.

Interactive debuggers are a great way to waste a ton of time and get absolutely nowhere. They do have their uses but those are not all that common. The biggest usecase for me for GDB has been to inspect stacktraces, having a good mental model of the software you are working on is usually enough to tell you exactly what went wrong if you know where it went wrong.

Lots of people spend way too much time debugging code instead of thinking about it before writing.

Oh, and testing >> debugging.

Nitpicking a bit here but theres nothing wrong with printf debugging. Its immensely helpful to debug concurrent programs where stopping one part would mess up the state and maybe even avoid the bug you were trying to reproduce.

As for tooling, I really love AI coding. My workflow is pasting interfaces in ChatGPT and then just copy pasting stuff back. I usually write the glue code by hand. I also define the test cases and have AI take over those laborious bits. I love solving problems and I genuinely hate typing :)

Yeah, I'm not putting stock in this at all. The methodology of this survey seems dubious.

I've tried the interactive debuggers but I'm yet to find a situation where they worked better than just printing. I use an interactive console to test what stuff does, but inline in the app I've never had anything that printing wasn't the straightforward fast solution.

The old fogeys don't rely on printf because they can't use a debugger, but because a debugger stops the entire program and requires you to go step by step.

Printf gives you an entire trace or log you can glance at, giving you a bird's eye view of entire processes.

The other thing they do is conveniently not mention all the negative stuff about AI that the source article mentions, they only report on the portion of content from the source that's in any way positive of AI.

And of course, its an article based on a source article based on a survey (of a single company), with the source article written by a "content marketing manager", and the raw data of the survey isn't released/published, only some marketing summary of what the results (supposedly) were. Very trustworthy.

>It's amazing how everyone can massage numbers when they're trying to sell something.

More like massage words.

Like one fairly smart person told me- Doing math with words is pointless.

I've used the 200 bucks Claude Code account, and while that's pretty great, it's definitely not "fire and forget". It often requires a completely different skill set. You basically need to develop a feeling for what the agent can do automatically and what needs extra prompting.

If you want a perfect investment imho, get Supermaven. It's autocomplete is 99% perfect.

For autocompletion I suggest giving Zed's own offering another try. It's guesses somewhat correctly at least half the time and is very convenient when it get's what you're trying to do durin a refactor and suggest valid multiline changes. It's still rather annoying when it replaces regular autocompletion for import paths with it's plausible sounding guesses. Before that I was using Cody, but I don't think that they have free tier any more.

> immediately consumed the entire free tier's worth of credits in just a few minutes of normal coding

Happened to me too when I've tried to do a single edit with connected gemini pro. It somehow ran out of free quota even before finishing it.

I don’t think you should look at the Copilot autocomplete as anything other than autocomplete, as far as I know they use worse but faster models for that.

I’m not big on AI but I get occasional use out of Copilot with the chat / agent mode and Claude Sonnet 4.

> suggestions are proactively generated and count against your account credit even if not accepted

This is no longer the case. They only count if you accept them.

Ime their model is not great, but better than copilot (ime copilot is just too slow and disruptive).

Also the options you have for tab-complete, afaik, are zed/zeta, guthub copilot, supermaven, and local models. I dont think there are other providers right now, glad if I am wrong in that.

Agreed. If I use an LLM to generate fairly exhaustive unit tests of a trivial function just because I can, that doesn’t mean those lines are as useful as core complex business logic that it would almost certainly make subtle mistakes in.

> test fixtures

I'm curious- how does the AI know what you want?

Yes, but in my experience actually no. At least not with the bleeding edge models today. I've been able to get LLMs to write whole features to the point that I'm quite surprised at the result. Perhaps I'm talking to it right (the new "holding it right"?). I tend to begin asking for an empty application with the characteristics I want (CLI, has subcommands, ...) then I ask it to add a simple feature. Get that working then ask it to enhance functionality progressively, testing as we go. Then when functionality is working I ask for a refactor (often it puts 1500 loc in one file, for example), doc, improve help text, and so on. Basically the same way you'd manage a human.

I've also been close to astonished at the capability LLMs have to draw conclusions from very large complex codebases. For example I wanted to understand the details of a distributed replication mechanism in a project that is enormous. Pre-LLM I'd spent a couple of days crawling through the code using grep and perhaps IDE tools, making notes on paper. I'd probably have to run the code or instrument it with logging then look at the results in a test deployment. But I've found I can ask the LLM to take a look at the p2p code and tell me how it works. Then ask it how the peer set is managed. I can ask it if all reachable peers are known at all nodes. It's almost better than me at this, and it's what I've done for a living for 30 years. Certainly it's very good for very low cost and effort. While it's chugging I can think about higher order things.

I say all this as a massive AI skeptic dating back to the 1980s.

I find this half state kind of useless. If I have to know and understand the code being generated, it's easier to just write it myself. The AI tools can just spit out function names and tools I don't know off the top of my head, and the only way to check they are correct is to go look up the documentation, and at that point I've just done the hard work I wanted to avoid.

Feels like a similar situation to self driving where companies want to insist that you should be fully aware and ready to take over in an instant when things go wrong. That's just not how your brain works. You either want to fully disengage, or be actively doing the work.

It turns into a specification problem.

This, IMHO, is the critical point and why a lot of “deep” development work doesn’t benefit much from the current generation of AI tools.

Last week, I was dealing with some temporal data. I often find working in this area a little frustrating because you spend so much time dealing with the inherent traps and edge cases, so using an AI code generator is superficially attractive. However, the vast majority of my time wasn’t spent writing code, it was getting my head around what the various representations of certain time-based events in this system actually mean and what should happen when they interact. I probably wrote about 100 test cases next, each covering a distinct real world scenario, and working out how to parameterise them so the coverage was exhaustive for certain tricky interactions also required a bit of thought. Finally, I wrote the implementation of this algorithm that had a lot of essential complexity, which means code with lots of conditionals that needs to be crystal clear about why things are being done in a certain order and decisions made a certain way, so anyone reading it later has a fighting chance of understanding it. Which of those three stages would current AI tools really have helped with?

I find AI code generators can be quite helpful for low-level boilerplate stuff, where the required behaviour is obvious and the details tend to be a specific database schema or remote API spec. No doubt some applications consist almost entirely of this kind of code, and I can easily believe that people working on those find AI coding tools much more effective than I typically do. But as 'manoDev says in the parent comment, deeper work is often a specification problem. The valuable part is often figuring out the what and the why rather than the how, and so far that isn’t something AI has been very good at.

I was told that I wasnt using it enough by one arm of the company and that I was spending too much by another.

Meanwhile, try as I might I couldnt prevent it from being useless.

I know of no better metaphor than that of what it's like being a developer in 2025.

Is he using it to code or just chatting up his AI girlfriend.

Claude is making $72k a year for a consistent $300/day spend.

Disclaimer: not only am I a Fastly employee, I'm the Fastly employee quoted in the register article originally linked.

We'd argue our stance and focus are a bit different. Our stance is -- we help eng teams and the open web with whatever their preference for AI might be.

If you want to go all in and monetize AI traffic or sign an agreement with a foundation model for training, we have tools that can help you do that. If you want to go scorched-earth on AI bots that try to scrape your systems, we can help you do that too.

It's all about helping engineers tasked with tough problems get their jobs done, regardless of whether the flavor of the week is AI or something else.

I've also just turned off copilot now. I had several cases where bugs in the generated code slipped through and ended up deployed. Bugs I never would have written myself. Reviewing code properly is so much harder than writing it from scratch.

I'm in the same exact boat. I started with a lot of different tools but eventually went back to hand coding everything. When using tools like co-pilot I noticed I would ship a lot mode dumb mistakes. I even experimented with not even using a chat interface and it turns out that a lot of answers to problems are indeed found with a web search.

I'm in this boat. On top of this though, I genuinely think less of developers who rely on LLMs. I very seriously believe that they are beneath me.

By all means, if my goal is actually crafting anything.

Usually it isn't, though - I just want to pump out code changes ASAP (but not sooner).

The article is saying older devs vibe code: I think you misunderstood

> LLMs are tools. If you master your tools, you become more productive.

LLMs are unpredictable tools which change all the time, let’s not pretend otherwise. You can’t “master” them in the same way as previous tools. You can learn some tricks to trick them to be closer to what you want, and that’s about it.

Imagine if every time you did the exact same movement to hammer a nail, you had to check your work. Maybe this time it hammered it in perfectly, or maybe it smashed your finger, or maybe it only went half-way through. You could never develop muscle memory for such a tool. You could use it, sure, but never master it.

a crack pipe is a tool. if you master your tools you become more productive. i've had the same mixed experiences with crack cocaine that others have. some good, some not so great. But i've been writing a lot of crack-fueled PRs in the last weeks. some needed manual intervention. some were tedious to create. and some clearly saved me a lot of work.

better code == less code? In my mind that would be highly succinct purely functional code in scala or haskell. Those really allow you to "compress" code down to its barest essential domain logic? Perhaps the future is writing highly domain specific DSLs and have the LLM generate the runtime for it.

> as long as we manage trust, ethics, and transparency

Trust, ethics and transparency are at all time lows in every corner of every capitalist industry imo

I don't know how anyone could look at the current state of things and think AI isn't being positioned to gut employment across many industries

Naive, childish and short sighted mentality imo

Yes there is even a small business that champions an small LLM that is trained on the language Lsp, your code base, and your recent coding history (not necessarily commit, but any time one presses ctrl + s). How it works is essentially autocomplete. This functionality is packaged as an IDE plugin: TabNine

However now they try to sell subscriptions to LLMs.

Tabnine has been in the scene since at least 2018.

> I've also learned enough about LLMs that I know how to write that CLAUDE.md so it sticks to best practices.

Could you share some examples / tips about this?

Can I ask which agent/model you used? I'm similarly irritated with shell script coding, but find I have to make scripts fairly often. My experience using various models but latterly Claude Code has been quite different -- it churned out pretty much what I was looking for. Also old, fwiw. I'm older than all shells.

Validity of the sample size is not determined by its fraction of the whole population. I don't know the formulas and I'm not a statistician. Maybe someone can drop some citations.

I generally agree with this just from a perspective of personal sentiment, it does feel wrong.

But statistically speaking, at a 95% confidence level you'd be within a +/- 3.5% margin of error given the 791 sample size, irrespective of whether the population is 30k or 30M.

You should read more about statistical significance. Under some reasonable assumptions, you can confidently certain deduce things with small sample sizes.

From another perspective: we've deduced a lot of things about how atoms work without any given experiment inspecting more than an insignificant fraction of all atoms.

TL;DR: The population size (25e6 total devs, 1e80 atoms in observable universe) is almost entirely irrelevant to hypothesis testing.

> Assuming they're like me, the planning is mostly done by a human, while the coding is largely done by the AI

I've largely settled on the opposite. AI has become very good at planning what to do and explaining it in plain English, but its command of programming languages still leaves a lot to be desired.

They're also being really loose with the term "older developers" by describing it as anybody with more than ten years of experience.

When debugging, I'll coax the AI to determine what went wrong first - to my satisfaction - and have it go from there. Otherwise it's a descent into madness.

as a human being i take it as massive disrespect when another human developer makes me wipe up their AI diarrhea.

[dead]

I think what happened is that a lot of people started dismissing all LLM code creation as "vibe coding" because those people were anti-LLM, and so the term itself became an easy umbrella pejorative.

And then, more people saw these critics using "vibe coding" to refer to all LLM code creation, and naturally understood it to mean exactly that. Which means the recent articles we've seen about how good vibe coding starts with a requirements file, then tests that fail, then tests that pass, etc.

Like so many terms that started out being used pejoratively, vibe coding got reclaimed. And it just sounds cool.

Also because we don't really have any other good memorable term for describing code built entirely with LLM's from the ground up, separate from mere autocomplete AI or using LLM's to work on established codebases.

I have "vibe coded" a few internal tools now that are very low risk in terms of negative business impact but nonetheless valuable for our team's efficiency.

E.g one tool packages a debug build of an iOS simulator app with various metadata and uploads it to a specified location.

Another tool spits out my team's github velocity metrics.

These were relatively small scripting apps, that yes, I code reviewed and checked for security issues.

I don't see why this wouldn't be a valid professional tool? It's working well, saves me time, is fun, and safe (assuming proper code review, and LLM tool usage).

With these little scripts it creates it's actually pretty quick to validate their safety and efficacy. They're like validating NP problems.

My personal definition of "vibe coding" is when a developer delegates -- abdicates, really -- responsibility for understanding & testing what AI-generated code is doing and/or how that result is achieved. I consider it something that's separate from & inferior to using AI as a development tool.

Yeah, for some reason the term has been used interchangeably for a while, which is making it very hard to have a conversation about it since many people think vibe coding is just using AI to assist you.

From Karpathy's original post I understood it to be what you're describing. It is getting confusing.

I think there is actually a pressure to show thst you are using AI (stories of ceos firing employees who supposedly did not "embrace" ai). So people are over attributing to AI. Though originally VC was meant to be infinite monkey style button smashing, people are attributing to VC just to avoid the cross hairs.

I just got back into OpenSCAD after recently getting my first new 3D Printer in 10 years, so I basically had to relearn it. ChatGPT got the syntax wrong for the most basic of operations.

What sort of coding do you do? I can't imagine getting that level of code from AI

If you’re using ChatGPT directly for work then I believe that you are doing it so profoundly wrong, at this point, that you’re going to make really incorrect assumptions.

As we have all observed, the models get things wrong, and if you’re wrong 5% of the time, then ten edits in you’re at 60-40. So you need to run them in a loop where they’re constantly sanity checking themselves—-linting, styling, typing and testing. In other words; calling tools in a loop. Agents are so much better than any other approach it’s comical precisely because they’re scaffolding to let models self-correct.

This is likely somewhat domain-specific; I can’t imagine the models are that great at domains they haven’t seen much code in, so they probably suck at HFT infrastructure for example, though they are decent at reading docs by this point. There’s also a lot of skill in setting up the right documentation, testing structure, interfaces, etc etc etc to make the agents more reliable and productive (fringe benefit; your LLM-wielding colleagues actually write docs now, even if they’re full of em-dashes and emoji). You also need to be willing to let it write a bunch of code, look at it, work out why it’s structurally deficient, throw it away, and build the structure you want to guide it - but the typing is essentially free, so that’s tractable. Don’t view it as bad code, view it as a useful null result.

But if you’re not using Claude Code or Codex or Roo or relatives, you’re living in an entirely different world to the people who have gone messianic about these things.

How would you know?

Haha right. I would imagine a "Senior Developer" who is super into AI assisted coding would be more likely to come across this survey and want to participate.

as a developer my first priority is whether the software works, not whether it is fast or easy to develop

> Anything that makes development faster or easier is going to be welcomed by a good developer.

I strongly disagree. Struggling with a problem creates expertise. Struggle is slow, and it's hard. Good developers welcome it.

these companies are destined to fail, just tell the truth about this dog shit technology and watch them flounder.

(Article was https://www.theregister.com/2025/08/28/older_developers_ai_c... when this was posted; we've since changed it. We also changed the title.)

I guess the article was vibe coded.