What blows my mind is how long fable can go without human input. I played a few and all these games would need at least 2-3 hours of back and forward with a raw version of opus. Of course we have harness tools that can achieve nearly identical results, but it appears that fable has created its own meta-harness inside itself to expand a prompt into a full plan before execution which makes me think a lot of these improvements come from baking claude code data directly into the model and simply adapting the entire manual process of the "back and forward".

I guess that would explain why fable feels like it is able to guess what the human input will be before it ends turn and continue working.

I heard Claude Code used a different system prompt for Fable, with different instructions.

It would be pretty funny if most of the gains turned out to be from prompting. Probably not. But a little harness goes a long way!

https://news.ycombinator.com/item?id=48192383

Have you noticed Claude Code now gives a suggestion for what to reply with that you can autocomplete with tab? Wouldn't surprise me if Fable used that concept in some way, whether directly or indirectly in training.