Julian Neagu

One thing worth remembering is that private fields don't automatically make secrets safe. # stops outside code from reading the field, but API keys still need proper handling outside the class too.

The difficulty jump is an interesting one. Same move count doesn't mean the same experience. I like that you looked at how the taps interact instead of just adding easier levels.

I've been building small tools with AI, and this really hits home. I don't need to know every detail of a new stack, but I do need to know enough to spot bad decisions before they become bigger problems.

The rising validation loss is the part I'd dig into next. Training loss looks nice, but it doesn't mean much if the model gets worse on unseen cases. I'd try fewer steps and compare the same test set.

Love that the agent learned to leave room for a rider after getting it wrong the first time. That’s the kind of small fix worth adding to the rules so the next run gets it right.

The part about cached UI vs actual offline functionality is important. I've seen how easy it is to assume offline support means everything still works. Being clear about what needs a connection saves users a lot of confusion.

I've seen the same thing with small tools. It's easy to focus on hiding the logic, but trusting a price from the browser is the bigger mistake. Always let the server have the final say.

The bit about asking before recommending really hits home. I’ve seen AI tools make confident guesses when one simple question would’ve saved everyone time. Getting the right context first makes all the difference.

The fallback point hits home. Most apps are built for the happy path, then one service goes down and everything starts piling up. Returning something useful beats showing a 500.

Love the focus on understanding existing apps, not just generating new ones. That’s where AI agents could save devs a ton of time, especially when the project has years of logic behind it.