Harry Sild

@luke_pighetti @jezell Also CSS is an actual styling system albeit bloated. A system has patterns that can be consistently applied thus learned by AI. Flutter has retarded bespoke styling props on widgets. Not much of a system here to use. It's more like AI is raw pattern matching from training data.

@jezell also so much of the styling is implicit unlike react native / css where all the styling is explicit

can't help but wonder if flutter is at a massive disadvantage because token vs semantic density is so low

@luke_pighetti It is at a major disadvantage imo. And if you try to get AI to write more compact functional code and use Mix, even with skills/guardrails the overall OOP "Enterprise" shit in the training data still regularly overwhelms whatever guardrails you set up.

@csells I mean yes you can give them such "personalities" but it's much more trickier to ensure that you don't end up with an elaborate theatre act from your crew putting up a play.

@csells I have done something similar for my openclaw setup. But also trying to tune their personalities from being superficial "yes-man" sycophants to something more like pragmatists that actually disagree when something is iffy. But still having a level of disagree but commit attitude.

I'm busy bootstrapping a startup with a human co-founder, so we have the CEO and the CTO roles covered. But I need the rest! So I asked Claude to install the knowledge worker plug-ins and used the following prompt to fill in the rest of the company: "Let's create the employees of the company each as a sub-agent for each of the skills, e.g. UX designer, HR, finance, etc., with a name, e.g. Fred the CFO, Harriet from HR, etc. and a unique personality perfectly suited for their role in the company."

@jezell @mraleph Even then it can go off the rails at times. Like our project has a specific custom analyzer plugin, custom skills, tuned AGENTS.md but it still struggles with functional / signals approach. Thanks ex-android devs for fucking up the LLM training data with OOP cruft.

codex 5.3 on max high is the first model I could really use for anything non trivial. Any other model just wasn’t helping me for most things. Still has to be told how to do things and be given small tasks to get good results. It’s much more like pair programming or explaining to a junior dev how you want things done. Very good at refactoring, testing, and grunt work. Good agents.md file can keep it aligned to your preferences and the best practices for your project (stuff like make sure to run dart format / dart analyze, always validate with unit tests, etc. saves manual iteration and helps make up for the waiting). Definitely things that aren’t common in the training set and it doesn’t have examples of it can be hit or miss unless you give detailed instructions and tell it ahead of time how it needs to be done.

I really wish I was in living in the same parallel universe as AI slop wankers and could just spawn a freaking agent per item of my backlog... And not have to rm -rf the produced garbage later.

@mraleph Since some models like Codex on extra high mode are so slow it works out pretty ok. And even failures are rarely total failures. Being systematic about the process still yields something that moves the needle. This approach though can be mentally exhausting due to multitasking.

@mraleph Oh definitely can take longer for a singular task. But to start seeing real gains objectively is to have like 4-10+ of different tasks being tackled at once in separate AI instances via git worktrees or just have multiple instances of the same repo to avoid potential conflicts.

@mraleph But getting AI to tackle hard stuff in Dart is a surprising struggle vs Elixir, Rust or Zig. I've queried all top LLMs why is that and they all invariably point the finger at the massive Java 8 style "Enterprise (tm)" Dart code in the corpus of knowledge overpowering all else.

@mraleph Funnily enough Elixir author had the same vibe post not long ago and the community superstars managed to get better results. Not saying skill issue, but there definitely is more to getting stuff out of AI than just asking for the hard problems. x.com/i/status/20198…

@mraleph @bernaferrari The AI otherwise will inevitably try to cook the tests or reward seek the actual code being tested towards passing the test. Which can result it making the code crappier in other areas. Also BDD / Gherkin has proven surprisingly effective. Especially with the "outside" test suite

@mraleph @bernaferrari The testing story is and has been a constant annoyance with the models. Some companies (mine included) have started putting the test suite in a separate repo where the AI can't see or modify them and they get served as "external scenarios" to pass.

@mraleph @bernaferrari Codex is much better at following instructions and sticking with them past the first couple of messages / tool calls. Codex is also my go to for code reviews (have way less false positives when its code review gets challenged than Claude). But Codex also lacks "creativeness".

@bernaferrari Codex is the only one I have not taken for a good spin so far. It's on my to-do list to try it next.

@bernaferrari @mraleph No, Gemini sadly is not. And Google forcing the use of Gemini for work is such Google corpo bullshit. If Gemini would be competitive it would be chosen naturally by the users based on its merit alone.

@mraleph Gemini 3.1 on Droid or Antigravity are *MUCH* better than gemini cli. But gemini 3.1 is still not good for your work.

@shiweidu @definev2 Both Claude and especially Codex can be so incredibly inefficient at Dart code. Both try by default to write Java 8 Enterprise style OOP code. And it's incredibly hard to force the models out of that tainted local maxima. Like they will even argue against the analyzer warnings.

@definev2 Some things are simply not meant to be written by AI. Codex wrote thousands of lines of code, while I decided to delete them and rewrite everything from scratch. It only took a few hundred lines to get it all done. Claude failed in the given scenario 😂

450 lines of code achieve the functionality of thousands of lines of code, with 2x faster performance.💥

@imadetheseworks Well Heroine did get some raised eyebrows in PR reviews at first...

I'm the best at giving packages a stupid joke name and then later not being able to rename when they become sort of serious projects. Like what do you mean people at work now say "Tim's stupid sheets" and I can't even be angry 😡

@shiweidu I meant you should aim to write the analyzer plugin in a way that maximizes performance and not for clean code looks. Basically the same sort of systems programming approach that makes alien_signals fast.

@HarrySild The main reason is that there is very little documentation for the analyzer plugin. You can only read the built-in rules of the analyzer to understand how to write it, so I had to implement a prototype first to ensure it worked correctly.

Oref is about to release a new version, including a fix and a new feature. In the next version, Oref will include a built-in analyzer plugin to improve the developer experience. > Give me your feedback! I can process it quickly 💪 #FlutterDev #DartDev CC @HarrySild

@shiweidu I mean it's still at it's core a combination of lexer, parser and analyzer with the same sort of pitfalls these things have in any other language.

@shiweidu Very nice but definitely make sure to give it the alien_signals performance oriented systems programming touch as the analyzer is super sensitive to Big O oopsies and generally gets hairy in monorepos.

@jezell I find Dart also interesting in regards to the visuals topic. Dart might become one of the few languages that are expressive and ergonomic enough to represent visuals via core language. And here I don't mean the dogshit nested OOP spaghetti fest Flutter has taken.

@jezell Some Rust libraries like Slint bypass this issue wholly by using a custom markup/DSL for styling representation. Which I personally think is the correct approach for Rust. You get the flexibility and ergonomics for representing the visual part while having Rust handle the rest.

Rust is really coming along on the UI layer. Any UI framework that isn't looking at some kind of native cargo / rust integration is missing out on the future. With all the crates available these days, Rust interop needs to get the same level of love as C interop. x.com/adamdotdev/sta…