One feature from an ancient AI that was amazing?

I’ll tell you my favorite example.

I was building applications in Solidity with AI long before the “GPT got good” era. The reason it was possible came down to one feature: the KoboldAI user interface let you edit the AI’s outputs directly.

A lot of the features people now associate with agentic harnesses already existed there in some form, but output editing was the one that changed everything for me.

Why did it matter? Because I didn’t have to constantly re-prompt. I could simply rewrite or adjust the AI’s response into what I actually needed. That let me steer the model in a completely different way. Instead of fighting the prompt, I could correct the trajectory directly.

I would love to see this implemented with modern frontier AI systems. It would save tokens, improve steering accuracy, and give users much more control over the conversation.

I also noticed something important: when I made the small edits that were usually necessary, each successive output became much better. The model adapted to the correction, and the conversation improved without needing a whole new prompt every time.