Super Duper AI Programming For Young Ladies; Not Bad, But Not Psychic Either

Just like with regular programming, you leave the important stuff for last, and the first and foremost are your project requirements and tests.

This used to be the worst thing about programming, but now AI takes care of that too.

In late to 2026 most of what was trending in the top programming dump. Was agent skills, agent memory, agent skill creating and updating.

And that really told the whole story, the leading programming agents, could not write big programs, AI brains were too small.

The is a way to create good code, and that involves using AI, to build up the skills of your repository, and testing is they work.

For complex programs with a long todo, you need a test system that no only says fail, but not implemented yet.

The test system creates a todo, because the AI can’t remember and certainly not in long sessions or across them.

Superficially it appears that AI can’t program yet, and to fix it you need to imbue your blank repository with skills, including meta skills.

This means self testing, self improving, and self evaluating.

A dismayed or jealous programmer can check us here and say, in 2026 AI is still a lot of nothing.

But I saw what it can do, and the thing that is holding it back, is self evaluation and self repair, self refinement.

My critique of AI is that when you set this up, you can mostly cancel your expensive AI.

You are doing so much hand holding, that spicy local AI can do almost everything, and know to call for help.

You can make small programs, you can et away with medium sized programs, but driving larger autonomous undertakings like serial app creation.

That will require repository design, self testing, self improvement, and putting a lot of usage credits in self correcting.

I agree that AI is iffy, but, it can learn and build up its own library, and it will correct it self, and add more test to keep it self correct.

It will expand its library of lessons, and continue advancing.

My closing observation is this, the world’s leading coding agent, has set its way in stone, it produces results that work.

The second best AI, lets you decide how to drive your coding agent, they have not quite established a pattern that produces results.

And you do that by chatting with your AI about what you need, what you tried and where you failed, what you dislike, and don’t want to deal with.

But then you have to know to say, now create a repository for my program, with all these skills, related skill and bunch of meta maintenance skills.

And then the tests will show you, if that will work for you.

What you can’t do is create the perfect repository, too many instructions are as bad as no specific SKILLS.

You can’t create something out of nothing, or hope to use generic skills.

Like with standard programming, you have to to the boring stuff first, and then test to see if ti works.

Blind programming will fail, AI will do a bunch of things that deliver, but the product will be a mixed breed dog.

To be clear the thing that is different today, is that AI will do years of strong programming in just a few days.

The bottleneck of typing up the code has been erased, but you still need to make sure you set everything up correctly.

AI can’t do that for you, in a way that you need, maybe clear thinking will turn out to be a uniquely human trait.