The fact that you consider this “saving time” might show that you are not being diligent with your code.
So what if the BDD is done?
Read >> Write.
As the final step, you should be cleaning up AI slop, bloated code, useless dependencies, unnecessary tests—or the lack thereof—security gaps, etc. Either way, any fine-tuning should happen in the final step.
This review stage will still eat up any time gains.
Is the AI result good enough? fix it, or refeed to AI to fix it.
I agree with you, however:
> Read >> Write.
is what a lot of AI agent zealots see as no longer a thing. I have had multiple persons tell me that AI will just have to get better in reading sloppy code and humans should no longer having to be able to do it.
Those zealots, gosh.
Nonsense. When AI gets stuck in a suboptimal topology, it’s the human who nudges it out.
How will you maintain code if you’re not even able to read it?
Using AI to read introduces noise with every iteration. Did you see that trend of people asking AI to reproduce someone self over 70 times? The noise adds up—and the original meaning shifts.
AI is non deterministic.