I had never used Typescript, I still don't even know the syntax as I haven't looked at a single line of code or text. I've done everything through claude code.

I'm pretty surprised how far along I am after a week. I couldn't have done this in 3 months without Claude Code.

Obviously it's not perfect, there is a lot to still do. Obviously there are normal bugs from LLMs being generally rough, but Opus 4.5 has done a pretty outstanding job in my mind given the constraints.

it's always funny, i think the opposite. I use a massive CLAUDE.md file, but it's targetted towards very specific details of what to do, and what not to do.

I have a full system of agents, hooks, skills, and commands, and it all works for me quite well.

I believe is massive context, but targetted context. It has to be valuable, and important.

My agents are large. My skills are large. Etc etc.

If there was one personal taste, this is achievable.

There is not. There are a variety of personal tastes and they all conflict with each other.

Which is my fundamental problem with peer review. There should be one gatekeeper to a code base that enforces some style preference. Practically which style they choose matters very little compared to having one predictable and enforced style to follow.

To clarify, I am talking about more than surface level style.

A small example adding an item to a list, it might be preference whether you use “push” or “append”

Don’t care which, but it is maddening to have a unit of code that mixes them.

There are all these choices that are numerous that should be consistent, but too numerous to add to a standards document ahead of time.

None of us mature in the same directions at the same speed.

What is taken by one as taste may be for another a hard won lesson in neglect or negligence.

I kept explaining the reasoning behind my nits and built my own DB of issues, explaining why it was a problem and what to do to reference as I started to explain my points over and over to different colleagues.

I feel that just giving out what to change without context does little to teach the hard won lessons, and is everything that would annoy people who are trying to get a change in.

The thing that always bothers me is when the comments involve things that have nothing to do with the PR.

Like, they are reasonable ideas, but open up a new issue. If every reviewer wants to tackle large topics in the PR Review that have nothing to do with what specifically is happening, then it explodes and gets even harder for others to review now that we are changing things that have nothing to do with the change.

It can be tricky to make a change in a file that is considered to be rotten. Especially for high priority tickets. Yes, we want to clear all of this code up. But you have to be very careful not to punish people for trying.

I had to tell one guy to knock it off because I was the only person brave enough to touch certain files and he was quickly making me regret trying by marking blocking comments on things I already planned to address in the subsequent or following PR. But I have to keep the old and new stuff working at the same time, boyo, so tap the brakes.

You used to build it? It's wonderful, thank you for your work.

Before the internet we asked people around us in our sphere. If we wanted to know the answer to a question, we asked, they made up an answer, and we believed it and moved on.

Then the internet came, and we asked the internet. The internet wasn't correct, but it was a far higher % correct than asking a random person who was near you.

Now AI comes. It isn't correct, but it's far higher % correct than asking a random person near you, and often asking the internet which is a random blog page which is another random person who may or may not have done any research to come up with an answer.

The idea that any of this needs to be 100% correct is weird to me. I lived a long period in my life where everyone accepted what a random person near them said, and we all believed it.

If you are asking random people, then your approach is incorrect. You should be asking the domain experts. Not gonna ask my wife about video games. Not gonna ask my dad about computer programming.

There, I've shaved a ton of the spread off of your argument. Possibly enough to moot the value of the AI, depending on the domain.

This all assumes you have experts that you can talk to. But they might be difficult to find or expensive to hire. You wouldn't want to waste your lawyer's time on trivia.

That is why experts often publish books and articles, which is then corrected by other experts (or random people if it’s a typo). I’ve read a lot of books and I haven’t met any of their authors. But I’ve still learned stuff.

Yep. At that point you're doing research, and become familiar enough with the literature to know what's right is work.

Much like with Wikipedia, using AI to start on this journey (rather than blindly using quick answers) makes a lot of sense.

Most of us don't have the domain expert in range, particularly pre-internet. I was merely suggesting you ask the most expert of the domain you have access to and work your way up the tree of knowledge from there.

However, the sibling commenter about books, journals, etc., is also an excellent suggestion.

Before the internet, I didn't have the phone number of domain experts to just call and ask these questions. perhaps you did. For a lot of us, it was an entirely foreign experience to have domain experts at your finger tips.

Didn’t you have books? And teachers?

How is an LLM making stochastic inferences based on aggregations of random blog pages more likely to be correct than looking things up on decidedly non-random blog pages written by people with relevant domain knowledge?

Is the above comment a genuine question? I’m concerned it’s a rhetorical question that isn’t really getting to the heart of the matter; namely, what is the empirical performance? One’s ability to explain said performance doesn’t always keep up.

How about we pick an LLM evaluation and get specific? They have strengths and weaknesses. Some do outperform humans in certain areas.

Often I see people latching on to some reason that “proves” to them “LLMs cannot do X”. Stop and think about how powerful such a claim has to be. Such claims are masquerading as impossibility proofs.

Cognitive dissonance is a powerful force. Hold your claims lightly.

There are often misunderstandings here on HN about the kinds of things transformer based models can learn. Many people use the phrase “stochastic parrots” derisively; most of the time I think these folks are getting it badly wrong. A careful reading of the original paper is essential, not to mention follow up work.

I'm not making a blanket statement against LLMs for all use cases. I'm certain that LLMs are, for example, much more performant at indexing already-curated documents and locating information within them than humans operating manually are.

What I'm skeptical about isn't LLMs as a utilitarian tool to enhance productivity in specific use cases, but rather treating LLMs as sources of information in their own right, especially given their defining characteristic of generating novel text through stochastic inference.

I'm 100% behind RAG powering the search engines of the future. Using LLMs to find reliable sources within the vast ocean of dubious information on the modern internet? Perfect -- ChatGPT, find me those detailed blog posts by people competent in the problem domain. Asking LLMs to come up with their own answers to questions? No thanks. That's just an even worse version of "ask a random person to make up an answer on the spot".

of course there are problems, there always will be. no solution will ever remove them, but the trade offs seem massively better.

Yeah, GPT 5 got into death loops faster than any other LLM, and I stopped using it for anything more than UI prototypes.

Neat idea!

Thank you :)

Love how you experimented, you are a creative thinker.

Haha, thank you! I just like to build stuff.