Uploaded September 2026 | Updated November 2026, 3 days ago
In the last episode, AI helped me make a huge change to the internal structure of Tsugi’s lesson system — and when it was over, the code looked amazing, but I wasn’t sure I trusted it.
So this time, I don’t start over.
I stay in the branch.
I trace old production paths. I ask nasty questions about edge cases. I let CodeRabbit criticize the changes. I fix little things. I build new things. And every time the branch survives another test, I understand it a little better and trust it a little more.
Along the way, we clean up the new Lessons V2 authoring model, simplify web-link subtypes, improve icon handling, and eventually take a big step: building a new IMS Common Cartridge export directly from the new canonical lesson structure.
By the end, something important has changed. I’m no longer wondering whether I should throw the AI-generated work away. I’m starting to see a path toward putting it into production. The more I work in the branch and test it through real development, the more confidence I gain in the whole thing.
This is less about “AI wrote some code” and more about the harder question:
How does an experienced programmer decide when AI-generated code is actually trustworthy?
Topics include AI-assisted software development, production code risk, code review, testing, refactoring, Cursor, CodeRabbit, Tsugi Lessons V2, IMS Common Cartridge, LMS import/export, and the human responsibility for AI-generated code.
In the last episode, AI helped me make a huge change to the internal structure of Tsugi’s lesson system — and when it was over, the code looked amazing, but I wasn’t sure I trusted it.
So this time, I don’t start over.
I stay in the branch.
I trace old production paths. I ask nasty questions about edge cases. I let CodeRabbit criticize the changes. I fix little things. I build new things. And every time the branch survives another test, I understand it a little better and trust it a little more.
Along the way, we clean up the new Lessons V2 authoring model, simplify web-link subtypes, improve icon handling, and eventually take a big step: building a new IMS Common Cartridge export directly from the new canonical lesson structure.
By the end, something important has changed. I’m no longer wondering whether I should throw the AI-generated work away. I’m starting to see a path toward putting it into production. The more I work in the branch and test it through real development, the more confidence I gain in the whole thing.
This is less about “AI wrote some code” and more about the harder question:
How does an experienced programmer decide when AI-generated code is actually trustworthy?
Topics include AI-assisted software development, production code risk, code review, testing, refactoring, Cursor, CodeRabbit, Tsugi Lessons V2, IMS Common Cartridge, LMS import/export, and the human responsibility for AI-generated code.










