Comment by TheAceOfHearts | Hacker News Reader

Comment by TheAceOfHearts | Hacker News Reader

Here's my short review after playing around with Firebase Studio for ~30 minutes. First of all, I had to turn off Firefox Enhanced Tracking Protection because otherwise projects wouldn't load.

I gave it the following initial prompt:

> An app where you input a question, then flip some coins to generate an I Ching prediction, and it generates a prediction / fortune for you. Then this combination of results can be fed to Gemini AI to produce a more detailed prediction text.

It generated something that looked fine. When I input a question and press the button nothing happened. After asking it to fix the problem multiple times and having it fail, I looked at the browser console to figure out the errors it was getting. Then I copied those errors and told it to fix them. After a few iterations, it solved every error and would generate a result. It completely forgot the part where you are supposed to flip coins before getting a hexagram to generate a fortune. After a bit of prompting, I was able to get it to display the hexagram and input question. However, sometimes it becomes confused about which hexagram was generated.

Overall, my impression is that these tools are still in the toy novelty stage rather than something you'd want to use for anything important.

Here is a screenshot of the app output for the question: Will Hacker News like my vibe coded oracle? [0] As you can see, it says that the generated hexagram is 24 or 41, but in the fortune text below it says 11.

[0] https://files.catbox.moe/i8t7rw.png

I built a complete working application (errortexts.com) using an AI tool, so I have a little insight on this.

At first, the product I was using (lovable.dev) seemed to me exactly as you described. I gave it a basic app outline and hit run, and it produced something that superficially looked right but did nothing.

So I asked some other people for advice, and they said you have to hold its hand and go step by step. So I did.

I told it, give me a landing page that matches [product description], but implement nothing else. Then, ok, let's set up auth - add a sign in and sign up dialog. Then, ok, let's create a user account page. Bit by bit.

It succeeded wildly. I was able to build the whole thing in 3 days. I'm not capable of that on my own, it would have taken me 3 weeks. Sometimes the AI got stuck and I had to manually go in and accomplish what I wanted. It took over 100 steps to complete the product, and probably around 10-20 times I had to revert its changes and give it more specific instructions. I had to check its work at every iteration, just like with a junior developer.

But it worked. And it's going to get better. Would I use this for "something important"? Depends how you define that. I used it to build a working product. Would I start letting it modify an existing mature codebase willy-nilly? No, probably not. Would I let it write cryptographic logic or trust that it wrote bulletproof code from a security standpoint in a sensitive context? No.

But for a simple application, it was an incredibly powerful tool. Especially for something that didn't even exist just 2 years ago. Give this a decade and it's going to change all our careers even more than it already has.

Can you... provide the prompts? Because I have a hard time believing this.

I have tried the hand holding approach with Cursor. It doesn't work for me. I have to constantly correct and over correct. Getting auth working sounds insane to me.

Having just figured auth recently for an app I’m building, I’ll echo that sentiment. Maybe lovable has auth components pre-built and well integrated

Yes, it uses Supabase, doesn’t roll its own

What exactly surprises you? You should try Gemini 2.5 Pro in Google AI Studio. Set the temperature to around 0.3, and in the system prompt, tell it to only edit exactly what you ask for and nothing else.

This model works really well. For example, simple things like [1], [2], and [3] can apparently be generated with just a couple of prompts.

[1] https://koreanrandom.com/en/games/through-the-space/

[2] https://koreanrandom.com/en/games/tanchiki/

[3] https://koreanrandom.com/en/games/sharik/

According to the author, these were made with Gemini 2.5 Pro without any manual coding by a human.

Cursor isn't as powerful as Gemini in AI Studio because AI Studio gives you full control over the model's settings and how it processes code.

Plus, the massive 1 million token context window is incredibly helpful for working with large codebases. You can use tools like code2prompt and repomix to feed all the necessary context into AI Studio from the clipboard for those projects.

I have had Claude Code build authentication before successfully, following a pattern approximately like this: https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/

I just got auth working with JWT through total vibecoding, using Claude+RooCode. Other bits of the app I needed a couple of tries, but auth worked immediately. I guess these models have seen express + node + JWT a million times.

I don't have access to the precise prompt, but I told it something like "only implement basic authentication based on JWT, using just email + password.", then asked it to add a simple registration form, then the password reset flow... step by step, but with little guidance.

At every prompt I review the changes on git, and commit.

I love the Java code sample:

```

// It works, but we have no idea how to send an HTTP request in java.

// If you do, send us a nice example to support@errortexts.com.

// You just need to POST JSON with the keys `api_key` and optionally `message`.

```

That's brilliant. But it does make me wonder why the LLM couldn't have provided you a suitable Java code example?

I wondered the same. Those code examples took an hour or so to figure out. It's a very basic exercise, sending an HTTP request, but doing it in nine languages, the majority of which you've never used, requires a little research. The llm can generate them all, but I had to run them locally to make sure they worked, and they oftentimes didn't. Most of them were easy and only took a minute or two to figure out, but after an hour and a half I gave up on Java. I literally could not figure it out. I'm sure I could have if I went long enough, but at some point you just have to cut your losses and decide it's not worth your time. This was not going to make or break my product offering, I'm not trying to learn Java, and I just wasn't interested in spending more time trying to figure it out. The llm was happy to generate examples, but they didn't work.

With these tools you have to talk to them as if you were talking to a knowledgeable, but a bit clueless junior developer. Sometimes it's almost as if you were coding it, just without typing the code.

I put your prompt into the vibe coding tool I'm working on (shameless plug).

The first version[0] looked good, but when I inspected it I found that it just picked an I Ching prediction at random on the back-end, instead of actually flipping coins.

I updated the prompt to the following:

> Create app where you input a question, then flip coins to generate an I Ching prediction (client-side). First show the standard I Ching prediction and it's hexagram, and then use AI to generate a fortune analysis based on the prediction and your initial question.

And the result was much more laborious[1] of a UI :shrug:

[0]https://iching.magicloops.app

[1]https://iching2.magicloops.app

Reply on news.ycombinator.com