Hacker News Reader

AI is excellent for tasks you know how to do, but can't be arsed to spend the time.

Example: I wanted a tool that notifies me of replies or upvotes to my recent Hacker News comments. Grok3 did it in 133 seconds with Think mode enabled. Total time including me giving it the example HTML as attachment and writing the specs + pasting the response to a file and running it? About 5 minutes.

I know perfectly well how to do it myself, but do I want to spend the hour or so to write all the boilerplate for managing state and grabbing HTML and figuring out the correct html elements to poll? Fuck no.

Yes I find this to be it's best use case. Unfortunately for anything I actually need help with the results are often terrible.

From my experience using ai, if you don't write a really precise description of your initial requirements in the initial prompt, and it it doesn't one-shot the answer, I don't bother asking it to fix the mistake.

Unless you're using a LLM with really long context, it's prone that some context loss will happen soner or later - when that happens, pointing errors that dropped out of the context will just result in repeated or garbage output.

This is why platforms(?) like Claude Code, Cursor and Windsurf are essential.

Claude Code forces you to create a CLAUDE.md to direct how it works, with Cursor you can (and should) write Cursor Rules.

The difference with a good spec + AI vs just vibe coding from scratch is like night and day.

I don‘t use these tools daily because I have a hard time to commit to these workflows. But do I need to do this setup once, once per project or every time for one off things I might code?

You really don't need these for tiny one-off scripts, but they're essential for larger projects where the whole application can't fit into the LLM context at once.

Basically they're just a markdown files where you write what the project is about, where everything is located and things like that. Pretty much something you'd need to have for another human to bring them up to speed.

An example snippet from the project I just happened to have open:

## Code Structure

- All import processors go under the `cmd/` directory - The internal/ directory contains common utility functions, use it when possible. Add to it when necessary. - Each processor has its own subdirectory and Go package

## Implementation Requirements

- Maintain consistent Go style and idiomatic patterns - Follow the existing architectural patterns - Each data source processor should be implemented as a separate command

This way the LLM (Cursor in this project) knows to, for example, check the internal/ directory for common utils. And if it finds duplication, it'll automatically generate any common functions in there.

It's a way to add guidelines/rails for projects, if you don't add anything the LLM will just pick whatever. It may even change style depending on what's being implemented. In this project the Goodreads processor was 100% different from the Steam processor. A human would've seen the similarities and followed the same style, but the LLM can't do that without help.

why would I pay for the advanced features when I haven't been impressed with the free features? in fact Claude 3.5, which is what is available, is a nearly worthless product, with value comparable to a free search engine, and not even a very good one. It is usually incorrect and frequently in subtle ways that will cost me a lot of time.

pro AI people sound like someone with an expensive addiction trying to justify it. the free product is bad, so I just need to pay to see the light?

Why would Anthropic let me use a model for free that is going to make me more skeptical of their paid offerings unless it is pretty similar to the paid ones and they think it's good?

Just read the manual and write the code yourself. These toys are a distraction.

Free version of Claude sucks because the prompts get too long very quickly. I have the Pro version and I have completed many tasks successfully.

(I know what I was doing and had a great understanding of what is going on).

Like many tools, there is some user skill required. Certainly there are situations where AI assistants won’t help much, but if every single attempt you’ve made to use an AI coding assistant has been “useless”, you are either working in a very niche area or, perhaps more likely, it is user error on your own part.

There are plenty of people who are way too high on the current abilities of AI, but I find that the “AI is useless crowd” to be equally ridiculous.

It reminds me of early in my career working in statistics where the company I joined out of grad school was justifiably looking to move out of SAS and start working in R and Python. Many were enthusiastic about the change and quickly saw the benefit, but there were some who were too entrenched in their previous way of working, and insisted that there was no benefit to changing, they could do anything required in SAS, and stubbornly refused to entertain the idea that there was a benefit to be gained by learning a new skill.

You needn’t become an AI cultist. But with the number of people who are getting at least some benefit to using AI coding assistants, if you are finding it to be worthless in your personal experience, it may be worth stepping back and considering if there is something wrong with how you are trying to utilize it.

What I do is go back through the conversation history, select the response that has the somewhat working code, then submit a prompt with what I want changed. By selectively including context and adjusting temperature, top_p/k and sometimes swapping model or system prompt for a given query will give better results. Combine this with repeating the query multiple times with that same context, then select which result is the best and move on.

I like to think of AI as a force (or skill) multiplier. If you’re low skill, it doesn’t do much. The higher your skill, the more useful it is.

That hasn't been my experience, nor that of others using the AI.

It's a force constant, rather than a multiplier. If you're low skilled, asking it to do a low skilled task, it works fine. If you're high skilled, and asked it to do the low skilled task, you saved a tiny bit of time (less than that of the low skilled person).

But it cannot do a high skilled task (at least, not right now). It can pretend, which can lead the low skilled person astray (but not the high skilled person).

Therefore, all AI does is raise the floor of what is achievable by the laymen, rather than multiply the productivity of a high skilled programmer.

You could also be a mixed skilled developer. Good at regular code, architecture, and algorithms but not as familiar with a given UI framework. Having the LLM generate the html and css for a given layout description saves a lot of time looking through the docs.

Does it really? The thing is that there’s a domain model beneath each kind of library and if they solve the same problem, you will find that they will generally follows the same pattern.

Let’s take the web. React, Svelte, Angular, Alpine.js, all have the same problems they’re solving. Binding some kind of state to a template, handling form interactions, decompose the page into “components”,… once you got the gist of that, it’s pretty easy to learn. And if you care about your code being correct, you still have to learn the framework to avoid pitfalls.

Same things with 3D engines, audio frameworks, physic engines, math packages,…

Using myself as an example -- I'm a long time C programmer (occasionally in a professional setting, mostly personal or as a side-item on my primary professional duties). I've picked up other languages through the years, had to deliver a web based application a few years ago so I did a deep-dive into html5, css3, and javascript. Now javascript has evolved since then, and I lost a bit of what I learned.

So now I want to do a new web application -- If I fall back on my C roots, my Javascript looks a lot like C. Example: adding an item to an array. The C style in Javascript would be to track the length of the array in another variable "len", and do something like myarray[len++] = new_value;

I can feed this into an LLM, or even say "Give me the syntax to add a value to an array", and it gives me "myArray.push(newValue)", which reminds me that "Oh yeah, I'm dealing with a functional/object oriented language, I can do stuff like this". And it reminds me that camelCase is preferred in Javascript. (of course, this isn't the real situation I've run into, just a simplified example -- but I really don't have all the method names memorized for each data type. So in that manner it is useful to get more concise (and proper) code.

I'm sure this is valuable for you, but here is my point of view.

I've worked professionally in many languages like Perl, Python, Kotlin, C# and dabbled in Common Lisp, Prolog, Clojure, and other exotics one. Whenever I forgot the exact syntax (like the loop DSL in CL), I know that there is a specific page in the docs that details all of these information. So just a quick keyword in a search engine and I will find it.

So when I come back in a language I haven't used in a while, I've got a few browser tabs opened. But that only lasts for a few days until I get back in the grove.

So for your specific example, my primary choice would have been the IDE autocompletion, then the MDN documentation for the array type. Barring that, I would have some books on the language opened (if it were some language that I'm learning).

LLMs will be best at HTML and CSS among programming languages because even if it's subtly wrong the browser will still render something.

Still, I'm not sure I need to spend $250/yr to occasionally generate some HTML that I could use a template to generate

My high skilled example from the other day: I wrote an algorithm out line by line in detail and had Claude turn it into avx2 intrinsics. It worked really well and i didn't have to dig through the manuals to remember how to get the lanes shuffled in the way I needed. Probably saved me 10 minutes but it would have been an annoying 10 minutes. :)

But that's a very low level task, where i had already decided on the "how it should be done" part. I find in general that for things that aren't really obvious, telling the llm _how_ to do it produces much more satisfactory results. I'm not quite sure how to measure that force multiplier - I can put in the same amount of work as a junior person and get better output?

> Rule number 1 and the only rule: You need to be a subject matter expert.

Strong disagree. I've been coding for 25+years but never on the front-end side. I couldn't write JS w/ pen & paper with a gun against my head. But I know what to ask and how to make sure a react component does what I want it to do, with these tools.

Or, in other words, you are a subject matter expert and are agreeing with the person you’re responding to. Quote the full argument (emphasis added):

> Rule number 1 and the only rule: You need to be a subject matter expert. Be it program logic or be it programming language.

You are a subject matter in program logic, just not programming language. You are supporting their point, not disagreeing.

I see where you're coming from. But later in the message they go > and if you do not understand the reason for the code and the programming language, you will take so much more time than if you did not even use the AI.

I guess that's the part I disagree on. The programming language is largely irrelevant now is what I'm seeing. And especially the "time to first result" is orders of magnitude smaller when using AI assistants than having to rtfm/google/so every little problem I encounter in an unfamiliar language.

And I am in no way stating that my code will match an expert in that field, of course.

How many thousands of lines does your AI-generated frontend code have?

Do you have to maintain the code?

> head. But I know what to ask and how to make sure a react component does what I want it to do

That is the subject matter expertise that many lack.