Coding agents create bad code, big code, many big files, and they prevent you from staying in control.

I have a very clear message, to share:

If you application cannot be divided into small commands, if your commands are not lightweight, and brilliant...

And your command assembly is not architecturally sound, abandon it all, and start anew.

It is not worthy of you, it will never grant you power.

Not only is fat code not a spell, but it has a terrible smell.

Take your time, experiment and learn, but also stay within the limits of a web page.

And before you generate your invention, ask AI, what would make it better.

What do wizards do, to stay true?

It should come as no surprise that the more you use the paid systems, the more you will be made to pay, certainly eventually.

And as night follows the day, the more you successfully create, the more power you will desire the larger you inventions will become.

Unsustainable is the keyword here, we have to attack the systems themselves.

That means, and you better grab onto something, that means today’s software development strategies are hostile to AI.

You can throw a hundred dollars a month at it, but that is feeding a fat man, you are just creating more problems.

So we need to rebuild software development, the name of the new game is flow based programming.

It is a standing wave, in harmony, betwixed many evils.

Example, you can’t hope to feed Un*x into an AI, but you sure as sunshine can feed in one command at a time.

And that is all this new invention is, the slaughter of Code Assistants and the Model Context Protocol.

Again, I repeat, if you keep throwing bread at a fat man, you are not fixing anything, get over it, move on.

Because without flow based programming you just end up with a mess, that the AI will forever try to guess.

Those of you who have legacy code, and have been receiving a fraction of the new AI power, know this.

To you I have a simple word of advice, that will de3stroy you if you don’t follow it.

Rewrite, or die.

AI+FBP, time to show that the old dogs, can learn new tricks faster than the pups - giddy up.

What does this translate to for the rest of us, you divide applicaions into small programs that snap together by cables.

These programs do not execute one after another, they listen for command packets.

Compress image, serve web page, convert blog post text to html.

They are not generic things, they are not meant to be reusabele, just connectable.

You spend an afternnon strategizing, about what commands must connect together to make your application.

And then, you feed the list of commands, one by one, into any AI, public or paid.

You get your programs, connect them, and have them route and execute packets to make you applicaion go.

Anything more than that, and your system is no longer servicable by AI.

AI does not understand today's applicaions, it guesses untill tests pass, if you are lucky enough to have them.

That is not how you use AI, that is how you waste its potential.

For those of you who are still confused about sending commands through pipes who are unsure where to start, start somewhere here:

"Hello everybody out there using minix - I'm doing a (free) operating system (just a hobby, won't be big and professional like gnu) for 386(486) AT clones. This has been brewing since April, and is starting to get ready. I'd like any feedback on things people like/dislike in Minix, as my OS resembles it somewhat (same physical layout of the file-system (due to practical reasons) among other things)." --Torvalds