My Fixes Became the Lesson: Teaching AI Better Coding Habits

The worst code is not the code that is clearly wrong. It is the code that is almost right.

Coding agents are painfully good at almost right.

They can write code that compiles, looks modern, and seems fine at first glance. Then you look closer: one line uses an anti-pattern. You ask the agent to clean it up, and it replaces that anti-pattern with another one. So you try again, more carefully, with the kind of prompt that should leave no room for misunderstanding.

It avoids the pattern you called out and finds a third way to be almost right.

I know. I did not believe it either.

At that point, I was so frustrated that all I could think was, "I need to make sure I never have to go through this again."

I needed the agent to get better. More skilled, really.
Wait.
Skill.
That is what I needed.

So I made the failure repeatable, wrote down the lesson behind the fix, and tested whether the agent stopped falling for it.

I will walk through that path with real generated code: the bad output, the better version, the skill, and the tests that proved it worked. The examples use Java streams and optionals, but the method works for any language, library, or coding habit your coding agent keeps getting wrong.

You will leave knowing how to turn frustration into a skill: capture the almost-right output, write the version you wanted in the first place, turn the difference into a rule, test the same task with and without the skill, and prove you can turn almost-right into right - and keep it that way.