Today, Gemini 2.5 pro blamed its tools for bad edits, then blamed me for a failed build in unchanged files. It is now truly equal to a human developer.

Tfw you realize your home PC is running more dev tools than your work box. .

Whenever I hear an OG dev talk about dropping a stack of punch cards I think, "Was that before boxes were invented?".

On the one hand, @cursor_ai is currently my favorite agentic coding environment. On the other hand, @intellijidea is just way easier for navigating, editing code, and viewing complete change sets. I guess I'm a 2-editor dev for this project?😅.

Any HTMX devs have thoughts about short-lived auth tokens for SSE (since it has to be in the path) vs WebSockets or fetch?.

I just added JaCoCo to my BDD project, just for curiosity. As it turns out, deleting unused classes really helps your test coverage.

My experience with web app BDD/TDD in JUnit:. - Write a web service spec.- An hour coding the glue.- 5 mins coding the feature. Writing JUnit web service tests is strange. but it's faster than external tools with a live app and DB. And I am VERY clear on the app's behavior. 😅.

So, I tried @windsurf_ai in IntelliJ today. Disappointing. The accept/reject markers wouldn't go away. It lost conversations. "Cascade base" butchered the code, so it couldn't even compile. Claude and GPT 4.1 struggled b/c they didn't see files. Maybe it's better in VS Code?.

It's a little baffling to me that we try to write unit tests for inputs and outputs only (no internals), but we're ok when front end tests know the app's exact HTML structure. Is there a better way to test the front end?.

I love it when things work out. At work I'm looking at a Java web app with separate front-end, back-end, and library repos, using MySQL, web services, and a website. This happens to be a very similar architecture and stack to the project I've been working on at home! 😁.

The agent generally did what I wanted, but with lots of hand holding. It didn't understand the game. It forgot how to build and run. It got distracted by small details and missed big ones. But it worked. In the next year or so they may move from "junior dev" to just "dev".

I asked for one feature at a time: "Make an empty app"; "add a character selection screen for 4 players"; "move the stats to a tooltip"; etc. Other than language I didn't give specify any code details, although I did say to make the stats configurable. 4/.

The project is a simple deck building game, and it is a complete, working app. It's not a brilliant concept, but the point was to get a working game quickly. It took about 10 hours across 3 days, and about 2 hours of that was spent generating game assets (images). 3/.

First thing, here's the repo: You're welcome to check out the code and tell me just how much you hate JavaFX, how bad the AI is at testing, or whatever your pet peeve is. :). 2/.

I decided to vibe code an entire app. Here's how it went!. 🧵.

Overall, I'd call this a successful attempt! I suspect in the end that this'll be far more effective than vibe coding. 😉 Thanks to @claude3_5sonnet for pairing with me, and @davefarley77 who unknowingly inspired this experiment!. 5/5.

The copilot also showed some initiative, with mixed results. It kept insisting that we output JSON instead of the single string value specified in the requirements. But it also refactored the behavior tests to a more readable pattern. 4/.

With behavior defined, the AI was enthusiastic about the red-green-refactor sequence. And it worked! We created a web service with basic CRUD, a bit of logic, and a robust set of tests. The first test took a couple of hours, but then we did four more in 15 minutes!. 4/.

I switched to BDD pair programming from scratch. It took a few attempts for the AI to discuss first without writing code, but then we started to flow. I had to write the behavior part of the test, because even after discussion the AI wrote more typical class/method tests. 3/.

The copilot is like a smart Jr dev who gets distracted by anything and focuses on irrelevant details. My first attempt was to test working code. The AI kept trying to mock very basic classes, but when I tried to correct it, the AI just doubled down. So I changed my approach. 2/.