Der neue Monatsrückblick für Mai und Juni 2026 ist online!

[Humidor]

This isn't for people who can't relate to this or are opposed to it. Therefore, I'm only asking those who use AI to create scripts and have had successes. Failures are part of the process, and that's where my question comes in, for sharing experiences.

My experience is based on Codex (Pro subscription). It works about 50% of the time, but I haven't yet found the chat startup script that makes it work reliably. Sometimes it doesn't work at all, sometimes it works immediately. Having Codex develop directly consumes the subscription very quickly, so I introduced Chatgpt. With it, the documentation on GitHub was generated, and the commands for Codex were developed. Ultimately, Chatgpt, due to internal loops and other issues, didn't recognize itself as the architect, and I've currently discontinued it.

What do you do when you have a task that needs to be implemented by AI?

I look forward to constructive input.

Anyone who can't accept this or has a negative view should please refrain from expressing their opinion here. This thread is pro-AI, and opponents are welcome to start their own thread.

With that in mind, let's discuss it.

[FernetMenta]

I have tried several AI coding agents. Until a couple of weeks ago I did the most of my work with github copilot inside VS Code. Then I got some promo credits for Kiro. Kiro is absolutely great, I loved it after the first few minutes I worked with it. It is more than only vibe coding because it has an additional spec/desing mode where you can do requirements, architecture and design upfront. I haven't used Kiro to develop iobroker scripts, yet but I wrote a complete adapter with it. I am sure it can do user scripts as well.

[OliverIO]

@Humidor

How do you use Codex? Via the web interface, the JavaScript adapter function, or the shell program? You also mentioned ChatGPT at one point.

When I use Codex via the shell program, I get excellent results. Very often, the output is right the first time—or requires at most one or two minor corrections.

With the JavaScript adapter function (does that even use Codex?), a relatively large prompt is transmitted alongside your actual instructions.

When using the web interface, Codex has to gather the necessary information first, which also consumes (input) tokens.

I would recommend using Codex via the shell and asking it to create a "skill" for ioBroker scripts.
It then saves this in the appropriate directory for future use.

If Codex does things you don't want it to, just tell it so or modify the skill (or have it modified).

The same applies if Codex implements things that are incompatible with ioBroker.

Over time, it causes fewer issues.
That said, my results are actually very good even without using a skill.

You mentioned that it consumes a lot of tokens—could you explain that in a bit more detail?
In my experience, consumption stays within reasonable limits.
It’s also important to ensure the context fill level doesn't drop too far below 50%; that’s when you start getting nonsensical or unnecessary results.

[Humidor]

I've only recently started using AI, and these initial scripts and context level settings don't mean much to me yet. This is likely a major area where I need to optimize. I'd like to learn more about this.

I'm using Codex and Chgpt in the app on my MacBook. I'm letting Codex directly apply changes to the Synology Docker ioBroker instance, directly in the live system (I also have a trial version).

Scripts usually work, but then it forgets how or which functions work, and I have to rework them. Should I create a plugin or something similar for this? Changes in VIS2 initially worked immediately, but now they don't work at all.