I'm hearing rumors that version 5.6-sol is going to be blazing fast. Is it true?

I wonder when Cursor's Composer 3 will be released?

That said, Codex's standard mode is slow, but switching to FAST mode burns through tokens in a flash—I wonder if there's a better way?

If you’re using Vibe Coding to build a browser game without thinking about component design, even adding a single button later can turn into a huge pain. If you don’t properly separate and abstract shared UI and logic into reusable components, even small changes start taking a huge amount of time once the product gets close to 90% complete.

I switched the model I use in Codex from 5.5 to 5.6. With 5.5, whenever a problem appeared on the screen, it kept trying to fix only the frontend. It would not even start checking the backend until I got fed up and told it, “Check whether the frontend and backend are consistent.” But 5.6 immediately identified the cause as “a mismatch between the backend API’s processing time and the frontend’s waiting time.” Impressive. This looks very promising.😆

When doing AI-assisted coding on a project where the backend and frontend are separated, the 90% problem becomes much heavier. Sometimes the root cause is on the backend side, but because the bug is visible on the screen, I end up asking the AI to keep fixing the frontend. Then the backend state or API problem remains unresolved, while only the frontend gets patched on the surface. Naturally, the problem never really gets fixed. First checking whether the problem might be on the backend side is a basic part of web development. But when using AI, it is surprisingly easy to just throw the visible problem at the AI, and that basic principle starts to break down. The 90% problem in AI-assisted coding isn’t just about the AI’s ability. It also depends heavily on how the developer isolates and diagnoses the problem.

I don’t work much with WordPress these days, but in the past, I worked on WordPress plugin development and theme development. Recently, I got curious and looked at the official WordPress plugin directory for the first time in a long time. Then I noticed that the main names looked almost the same as about 10 years ago. Web development has changed completely in these 10 years, but the WordPress world almost looks like time has stopped. Maybe this means that if you have a good plugin idea, this area could still be a pretty good target?

I tried using Codex Security on a setup where the backend and frontend are separated, and it pointed out several things I had never noticed before. And then it burned through my 100 USD plan quota in no time. This thing is amazing. Both the performance and the token consumption are incredible.

Embarrassingly, yesterday I discovered a Phaser-oriented skill in Codex’s “Game Studio” plugin. When I type `/phaser` before giving instructions, its ability to understand Phaser’s structure seems to improve. Yesterday I also ran into the 90% problem, and it just wouldn’t fix things the way I wanted. But after I noticed the `/phaser` slash command and put it at the beginning of my instruction, it fixed it in one shot. But well, this time it was only a frontend issue. When making a game with Phaser, if you want to save multiple users’ progress or add authentication, you will probably need to connect it with a backend. When the frontend and backend are separated like this, even with /phaser added, it still seems like handling that is going to be difficult.

Phaser Editor has evolved significantly with version 5. At this rate, I might not even need to use Tiled to create maps anymore...

At the finishing stage of a browser game, I sometimes find myself pausing over the question of what to do when a user opens the same game in multiple tabs. To deal with this, I had been using WebSocket so that if the same game was opened in multiple tabs, a warning would be shown and the game would be made unplayable. However, the other day, while casually looking through the MDN documentation, I learned that there is a feature called Web Locks. It is possible to detect multiple tabs playing the same game with WebSocket as well, but it requires writing code on both the frontend and the server side. Also, in my implementation, the server would detect it via WebSocket and then return a warning, so there was a slight delay in detection, and it didn’t always fully function as a guard at the exact moment the page was opened. In that sense, if the goal is simply to prevent multiple instances within the same browser, Web Locks is simpler in terms of code, and since it completes within the browser, the response is faster. Recently, I had been relying almost entirely on asking Codex to “make this,” but unless I wander around the internet and explore technologies like before Coding AI existed, I wouldn’t have been able to find Web Locks. Web development is not yet something that can be entirely left to AI.

I’ve been trying for a while to somehow automate pixel animation, and I’ve been experimenting with generating sprite sheets using GPT Image Gen 2. But in the end, they still only look like they might be usable for pixel animation, while in reality they aren’t usable at all... When I cut out each frame one by one and apply them frame by frame in Aseprite, parts that seem like they should be placed in the same position are sometimes shifted by just a few pixels. When that happens, the image falls apart when played as an animation, and it becomes unusable. So for now, even if I make this kind of sprite sheet, I can only use it as a rough reference, and I still have no choice but to build the pixels manually myself. I just hope the day comes soon when this can finally be automated...

Codex is acting weirdly smart today. Did they sneak 5.6 in there or what???

GLM’s Coding Plan feels a bit pricey. If it’s 64.8 USD/month, I feel like GPT’s 100 USD plan might be better for me, since it also gives me Pro mode and image generation...

When Codex modifies the frontend, it creates build files on its own, but for some reason, what it shows in the browser always seems to be the result of the previous build. Thanks to that, I’ve been struggling with the problem where even after modifying the frontend, I can’t actually confirm the result. So I wrote “Don’t access build files during development” in AGENTS.md, and the problem was solved in one shot.

During the Fable 5 trial, I was blown away seeing visually amazing games being mass-produced from just a single-line prompt. But if this many visually impressive games are mass-produced, doesn’t the beauty of visuals start to become meaningless? Big-hit games, games that go down in history, are often revolutionary not in their visuals, but in their game logic. Take Vampire Survivors, for example. If you judge it only by its looks, it looks like a game from 40 years ago. But its core gameplay structure — “even without attack controls toward enemies, you can still defeat them and grow your character” — was revolutionary. The more advanced AI visuals become, the more convinced I am that the value of basic game logic will rise instead.

I just lost two hours with Codex’s GPT-5.5 xhigh on a Phaser Scene launch that should’ve been solved by inserting a one-line function. What is going on? Is this a sign that GPT-5.6 is coming, or was my prompt just stupid?

AI coding and app development have already become normal, but the “90% problem” that people were talking about back in the GPT-4 days still remains. Especially when it’s combined not only with frontend, but also backend, authentication, containerization, and so on, the last 10% becomes a road of hardship. The closer it gets to completion, the harder it becomes to adjust the details. How is everyone overcoming this problem? Should we expect GPT-5.6 to solve it? 😅

After trying over and over again, it still seems difficult to create pixel art with gpt-image-2. Even when I ask Codex to “make it strictly on a 32×32 grid,” and then shrink the generated image down to 32px × 32px in Affinity to check it, the pixels are slightly off. They don’t snap precisely to the 1px grid. I also tried using the Canva plugin, but in that case, the quality of the artwork itself wasn’t very high. For now, if I want to use gpt-image-2 to create pixel animation, the most realistic workflow may be to generate an image first, bring it into Affinity or Photoshop, manually redraw it, and then import that asset into Aseprite. I thought pixel art would be one of the first things AI would automate, but apparently not. It looks like there are still things humans can do.😀