I second this. This* is the matter against which we form understanding. This here is the work at hand, our own notes, discussions we have with people, the silent walk where our brain kinda process errors and ideas .. it's always been like this since i was a kid, playing with construction toys. I never ever wanted somebody to play while I wait to evaluate if it fits my desires. Desires that often come from playing.

Outsourcing this to an LLM is similar to an airplane stall .. I just dip mentally. The stress goes away too, since I assume the LLM will get rid of the "problem" but I have no more incentives to think, create, solve anything.

Still blows my mind how different people approach some fields. I see people at work who are drooling about being able to have code made for them .. but I'm not in that group.

That's also how I feel.

I think you have every right to doubt those telling us that they run 5 agents to generate a new SAAS-product while they are sipping latté in a bar. To work like that I believe you'll have to let go of really digging into the code, which in my experience is needed if want good quality.

Yet I think coding agents can be quite a useful help for some of the trivial, but time consuming chores.

For instance I find them quite good at writing tests. I still have to tweak the tests and make sure that they do as they say, but overall the process is faster IMO.

They are also quite good at brute-forcing some issue with a certain configuration in a dark corner of your android manifest. Just know that they WILL find a solution even if there is none, so keep them on a leash!

Today I used Claude for bringing a project I abandoned 5 years ago up to speed. It's still at work in progress, but the task seemed insurmountable (in my limited spare time) without AI, now it feels like I'm half-way there in 2-3 hours.

>but the code to me is a forcing mechanism into ironing out the details, and I don't get that when I'm writing a specification.

This is so on point. The spec as code people try again and again. But reality always punches holes in their spec.

A spec that wasn't exercised in code, is like a drawing of a car, no matter how detailed that drawing is, you can't drive it, and it hides 90% of the complexity.

To me the value of LLMs is not so much in the code they write. They're usually to verbose, start building weird things when you don't constantly micromanage them.

But you can ask very broad questions, iteratively refine the answer, critique what you don't like. They're good as a sounding board.

I still do this, but when I'm reviewing what's been written and / or testing what's been built.

How I see it is we've reverted back to a heavier spec type approach, however the turn around time is so fast with agents that it still can feel very iterative simply because the cost of bailing on an approach is so minimal. I treat the spec (and tests when applicable) as the real work now. I front load as much as I can into the spec, but I also iterate constantly. I often completely bail on a feature or the overall approach to a feature as I discover (with the agent) that I'm just not happy with the gotchas that come to light.

AI agents to me are a tool. An accelerator. I think there are people who've figured out a more vibey approach that works for them, but for now at least, my approach is to review and think about everything we're producing, which forms my thoughts as we go.

I also second this. I find that I write better by hand, although I work on niche applications it’s not really standard crud or react apps. I use LLMs in the same way i used to used stack overflow, if I go much farther to automate my work than that I spend more time on cleanup compared to if I just write code myself.

Sometimes the AI does weird stuff too. I wrote a texture projection for a nonstandard geometric primitive, the projection used some math that was valid only for local regions… long story. Claude kept on wanting to rewrite the function to what it thought was correct (it was not) even when I directed to non related tasks. Super annoying. I ended up wrapping the function in comments telling it to f#=% off before it would leave it alone.

Exactly. 30 years ago a mathematician I knew said to me: "The one thing that you can say for programming is that it forces you to be precise."

We vibe around a lot in our heads and that's great. But it's really refreshing, every so often, to be where the rubber meets the road.

Historically software engineering has been seen as "assembly line" work by a lot of people (see all the efforts to outsource it through spec handoffs and waterfall through the years) but been implemented in practice as design-as-you-build (nobody anticipates all the questions or edge cases in advance, software specs are often an order of magnitude simpler than the actual number of branches in the code).

For mission-critical applications I wonder if making "writing the actual code" so much cheaper means that it would make more sense to do more formal design up front instead, when you no longer have a human directly in the loop during the writing of the code to think about those nasty pops-up-on-the-fly decisions.

> A lot of how I form my thoughts is driven by writing code, and seeing it on screen, running into its limitations.

Two principles I have held for many years which I believe are relevant both to your sentiment and this thread are reproduced below. Hopefully they help.

First:

  When making software, remember that it is a snapshot of 
  your understanding of the problem. It states to all, 
  including your future-self, your approach, clarity, and 
  appropriateness of the solution for the problem at hand. 
  Choose your statements wisely.

  Code answers what it does, how it does it, when it is used, 
  and who uses it. What it cannot answer is why it exists. 
  Comments accomplish this. If a developer cannot be bothered 
  with answering why the code exists, why bother to work with 
  them?

In 1987 when I first started coding, I would either write my first attempt in BASIC and see it was too slow and rewrite parts in assembly or I would know that I had to write what I wanted from the get go in assembly because the functionality wasn’t exposed at all in BASIC (using the second 64K of memory or using double hires graphics).

This past week, I spent a couple of days modifying a web solution written by someone else + converting it from a Terraform based deployment to CloudFormation using Codex - without looking at the code as someone who hasn’t done front in development in a decade - I verified the functionality.

More relevantly but related, I spent a couple of hours thinking through an architecture - cloud + an Amazon managed service + infrastructure as code + actual coding, diagramming it, labeling it , and thinking about the breakdown and phases to get it done. I put all of the requirements - that I would have done anyway - into a markdown file and told Claude and Codex to mark off items as I tested each item and summarize what it did.

Looking at the amount of work, between modifying the web front end and the new work, it would have taken two weeks with another developer helping me before AI based coding. It took me three or four days by myself.

The real kicker though is while it worked as expected for a couple of hundred documents, it fell completely to its knees when I threw 20x documents into the system. Before LLMs, this would have made me look completely incompetent telling the customer I now wasted two weeks worth of time and 2 other resources.

Now, I just went back to the literal drawing board, rearchitected it, did all of the things with code that the managed services abstracted away with a few tweaks, created a new mark down file and was done in a day. That rework would have taken me a week by itself. I knew the theory behind what the managed service was doing. But in practice I had never done it.

It’s been over a decade where I was responsable for a delivery that I could do by myself without delegating to other people or that was simple enough that I wouldn’t start with a design document for my own benefit. Now within the past year, I can take on larger projects by myself without the coordination/“mythical man Month” overhead.

I can also in a moment of exasperation say to Codex “what you did was an over complicated stupid mess, rethink your implementation from first principles” without getting reported to HR.

There is also a lot of nice to have gold plating that I will do now knowing that it will be a lot faster

That's because many developers are used to working like this.

With AI, the correct approach is to think more like a software architect.

Learning to plan things out in your head upfront without to figure things out while coding requires a mindset shift, but is important to work effectively with the new tools.

To some this comes naturally, for others it is very hard.

Using AI or writing your own code isn't an xor thing. You can still write the code but have a coding assistant or something an alt/cmd-tab away. I enjoy writing code, it relaxes me so that's what I do but when I need to look something up or i'm not clear on the syntax for some particular operation instead of tabbing to a browser and google.com I tab to the agent and ask it to take a look. For me, this is especially helpful for CSS and UI because I really suck at and dislike that part of development.

I also use these things to just plan out an approach. You can use plan mode for yourself to get an idea of the steps required and then ask the agent to write it to a file. Pull up the file and then go do it yourself.

> A lot of how I form my thoughts is driven by writing code, and seeing it on screen, running into its limitations.

I completely agree but my thought went to how we are supposed to estimate work just like that. Or worse, planning poker where I'm supposed to estimate work someone else does.

> A lot of how I form my thoughts is driven by writing code, and seeing it on screen, running into its limitations.

If you need that, don't use AI for it. What is it that you don't enjoy coding or think it's tangential to your thinking process? Maybe while you focus on the code have an agent build a testing pipeline, or deal with other parts of the system that is not very ergonomic or need some cleanup.

Any sufficiently detailed specification converges on code.

i go back and forth on this. when i'm working on something where the hard part is the actual algorithm, say custom scheduling logic or a non-trivial state machine, i need my hands in the code because the implementation is the thinking. but for anything where the complexity is in integration rather than logic, wiring up OAuth flows, writing CRUD endpoints, setting up CI pipelines, agents save me hours and the output is usually fine after one review pass. the "code as thought" argument is real but it applies to maybe 20% of what most of us ship day to day. the other 80% is plumbing where the bottleneck is knowing what to build, not how.

I couldn't agree more. It's often when you are in the depth of the details that I make important decisions on how to engineer the continuation.

Sounds like the coders equivalent of the Whorfian hypothesis.

I sometimes wonder if the economics of AI coding agents only work if you totally ignore all the positive externalities that come with writing code.

Is the entire AI bubble just the result of taking performance metrics like "lines of code written per day" to their logical extreme?

Software quality and productivity have always been notoriously difficult to measure. That problem never really got solved in a way that allowed non technical management to make really good decisions from the spreadsheet level of abstraction... but those are the same people driving adoption of all these AI tools.

Engineers sometimes do their jobs in spite of poor incentives, but we are eliminating that as an economic inefficiency.

I dunno. On the one hand, I keep hearing anecdata, including hackernews comments, friends, and coworkers, suggesting that AI-assisted coding is a literal game changer in terms of productivity, and if you call yourself a professional you'd better damn well lock the fuck in and learn the tools. At the extreme end this takes the form of, you're not a real engineer unless you use AI because real engineering is about using the optimal means to solve problems within time, scale, and budget constraints, and writing code by hand is now objectively suboptimal.

On the other hand, every time the matter is seriously empirically studied, it turns out that overall:

* productivity gains are very modest, if not negative

* there are considerable drawbacks, including most notably the brainrot effect

Furthermore, AI spend is NOT delivering the promised returns to the extent that we are now seeing reversals in the fortunes of AI stocks, up to and including freakin' NVIDIA, as customers cool on what's being offered.

So I'm supposed to be an empiricist about this, and yet I'm supposed to switch on the word of a "cool story bro" about how some guy built an app or added a feature the other day that he totally swears would have taken him weeks otherwise?

I'm like you. I use code as a part of my thought process for how to solve a problem. It's a notation for thought, much like mathematical or musical notation, not just an end product. "Programs must be written for people to read, and only incidentally for machines to execute." I've actually come to love documenting what I intend to do as I do it, esp. in the form of literate programming. It's like context engineering the intelligence I've got upstairs. Helps the old ADHD brain stay locked in on what needs to be done and why. Org-mode has been extremely helpful in general for collecting my scatterbrained thoughts. But when I want to experiment or prove out a new technique, I lean on working directly with code an awful lot.

I was just thinking this the other day after I did a coding screen and didn't do well. I know the script for the interviewee is your not suppsed to write any code until you talk through the whole thing, but I think i woukd have done better if I could have just wrote a bunch of throw away code to iterate on.

Are there still people under the impression that the correct way to use Stack Overflow all these years was to copy & paste without analyzing what the code did and making it fit for purpose?

If I have to say, we're just waiting for the AI concern caucus to get tired of performing for each other and justifying each other's inaction in other facets of their lives.

this is close to what i've landed on too. the pre-commit hook is non-negotiable. i've had Claude Code report "all checks pass" when there were 14 failing eslint rules. beyond the static analysis though, i keep hitting a harder problem: code that passes every lint rule, compiles clean, and greens the test suite but implements a subtly wrong interpretation of the spec. like an API handler that returns 200 with an empty array instead of 404, technically valid but semantically wrong. evaluating behavioural correctness against intent, not just syntax or type safety, is the gap nobody's really cracked yet. property-based testing helps but it still requires you to formalize the invariants upfront, which is often the hard part.

My setup has some of the things mentioned and I found that occasionally the LLM will lie that something passes, when it doesn't.

These kinda things aren’t really the issues I run into. Lack of clarity of thought, overly verbose code, needlessly defensive programming - the stuff that really rots a codebase. Honestly some of the above rules you have I’d want the LLM to ignore at the times if we’re going for maximum maintainability.

Except for dependency cruiser which I hadn't heard of, this is almost exactly what I've built up over the past few weeks.

For the pre-commit hook, I assume you run it on just the files changed?

> Custom script to ensure shared/util directories are not over stuffed (built this using dependency-cruiser as a library rather than an exec)

Would you share this?

It’s a solid post overall and even for people with a lot of experience there’s some good ideas in here. “Identify and mark functions that have a high security risk, such as authentication, authorization” is one such good idea - I take more time when the code is in these areas but an explicit marking system is a great suggestion. In addition to immediate review benefits, it means that future updates will have that context.

“Break things down” is something most of us do instinctively now but it’s something I see less experienced people fail at all the time.

Anyone who’s been a developer for more than 10 minutes knows that best practices are hard to always follow through on when there’s pressure to ship.

But there’s more time to do some of these other things if the actual coding time is trending toward zero.

And the importance of it can go up with AI systems because they do actually use the documentation you write as part of their context! Direct visible value can lead people to finally take more seriously things that previously felt like luxuries they didn’t have time for.

Again if you’ve been a developer for more than 10 minutes, you’ve had the discouraging experience of pain-stakingly writing very good documentation only for it to be ignored by the next guy. This isn’t how LLMs work. They read your docs.

We’re not „thinking with portals” about these things enough yet. Typically we’d want a detailed spec beforehand, as coding is expensive and time consuming, thus we want to make sure we’re coding the right thing. With AI though, coding is cheap. So let AI skip the spec and write the code badly. Then have it review the solution, build understanding, design a spec for better solution and have it write it again. Rinse and repeat as many times you need.

It’s also nothing new, as it’s basically Joe Armstrong's programming method. It’s just not prohibitively expensive for the first time in history.

Astronaut 1, AI-assisted developers: You mean, it's critical to plan and spec out what you want to write before you start in on code?

Astronaut 2, Tim Bryce: Always has been...

I’m finding it to be the opposite. I used to love writing everything by hand but now Claude is giving me the ability to focus more on architecture. I like just sitting down with my coffee and thinking about the next part of my project, how I’d like it to be written and Claude just fills it in for me. It makes mistakes at times but it also finds a lot of mine that I hadn’t even realized were in my code base.

A good bit of scaffolding and babysitting allows you to let the model run much faster and more efficiently. Building your tool faster. I don't code to code, I code to build something I want.

Each extra generation of AI produced crap AI consumes as training, the worse it gets. This has been mathematically proven.

They will also be reading all of the slop generated by the current and previous generations of LLMs

Pre commit hook is definitely necessary. One thing I’ve seen a lot with Opus recently is it lying that a new linter warning or error was there before it made a change. They’ve learned from us too well!

small agents.md files are worth it, at least for holding some basic information (look at build.md to read how to build, the file structure looks like so), rather than have whatever burn double the amount of tokens searching for whatever anyways.

Underground Resistance Aims To Sabotage AI With Poisoned Data

https://news.ycombinator.com/item?id=46827777

I’ll believe it when those same engineers fix CC’s awful performance (mostly kidding, though I do wonder why they can’t. Feels like it’s doable).

In reality that man is hoping to IPO in 6-12 months, if anyone is wondering why the “use claude or you’re left behind” is so heavy right now.

I'm finding Rust is perfect for me with LLMs.

I find rust generally easier to reason about, but can't stand writing it.

The compiler works well with LLMs plenty of good tooling and LSPs.

If I'm happy with the shape of the code and I usually write the function signatures/ Module APIs. And the compiler is happy with it compiling. Usually the errors if any are logical ones I should catch in reviews.

So I focus on function, compiler focuses on correctness and LLM just does the actual writing.

Do you think Rust will end up getting a boost from LLM adoption?

  > repo will always win over documentation

i think languages that allow expression of "this is deprecated, use x instead" will be usefull for that too

"Stack Overflow that reads your codebase" — perfect. But Stack Overflow is stateless. Agent sessions aren't.

One session's scaffold assumes one pattern. Second session scaffold contradicts it. You reviewed both in isolation. Both looked fine. Neither knows about the other.

Reviewing AI code per-session is like proofreading individual chapters of a novel nobody's reading front to back. Each chapter is fine. The plot makes no sense.

Every engineering org should be pleading devs to not let AI write code, period. They continue to routinely get stuff wrong and can't be trusted any further than you can throw them.

Same. Small units if work, iterate in it till it's right, commit it, push it, then do the next increment of work. It's how I've always worked like that, except now, I sometimes let someone else figure the exact API calls (I'm still learning react, but Claude helps get the basics in place for me). If the AI just keeps screwing up, I'll grab the wheel and do it myself. It sometimes helps me get things going, but it hasn't been a huge increase in productivity, but I'm not paying the bill so whatever.

Hi raphman,

i have written this text by myself except like 2 or 3 sentences which i iterated with an LLM to nail down flow and readability. I would interpret that as completely written by me.

> The suggestions you make are all sensible but maybe a little bit generic and obvious. Asking ChatGPT to generate advice on effectively writing quality code with AI generates a lot of similar suggestions (albeit less well written).

Before i wrote this text, i also asked Gemini Deep Research but for me the results where too technical and not structural or high level as i describe them here. Hence the blogpost to share what i have found works best.

> If not, I'd suggest to augment the post with practical examples or anecdotal experience. At the moment, the target group seems to be novice programmers rather than the typical HN reader.

I have pondered the idea and also wrote a few anecdotal experiences but i deleted them again because i think it is hard to nail the right balance down and it is also highly depended on the project, what renders examples a bit useless.

And i also kind of like the short and lean nature of it the last few days when i worked on the blogpost. I might will make a few more blogposts about that, that will expand a few points.

Thank you for your feedback!

That's the opposite. I've never read and re-read code more than i do today. The new hires generate 50 more code than they use to, and you _have_ to check it or have compounding production issues (been there, done that). And the errors can now be anywhere, when before you more or less knew what the person writing code is thinking and can understand why some errors are made. LLMs errors could hide _anywhere_, so you have to check it all.

They serve as guardrails for agents to not do stupid things.

If your goal is for AI to write code that works, is maintainable and extensible, you have to include as many deterministic guardrails as possible.

Man, you are really missing out of the biggest revolution of my life.

This is the opinion of someone who has not tried to use Claude Code, in a brand new project with full permissions enabled, and with a model from the last 3 months.

This take is pretty uncharitable. I write high quality code, but also there's a bunch of code that could be useful, but that I don't write because it's not worth the effort. AI unlocks a lot of value in that way. And if there's one thing my 25 years as a software engineer has taught me is that while code quality and especially system architecture matter a lot, being super precious about every line of code really does not.

Don't get me wrong, I do think AI coding is pretty dangerous for those without the right expertise to harness it with the right guardrails, and I'm really worried about what it will mean for open source and SWE hiring, but I do think refusing to use AI at this point is a bit like the assembly programmer saying they'll never learn C.

Can you be specific? You didn't provide any constructive feedback, whatsoever.

Look up luddites on Wikipedia, might be too deep to see the similarities though.