Ask HN: How to make AI better at writing prose?

[remywang]

Current models are still notoriously bad at writing prose. It's not just the easily recognizable tics like "load-bearing" and "genuinely". The text is just sloppy, pompous, and full of air.

I've tried different approaches like feeding it examples of my own writing, giving it a style guide, and iterating a few editing sessions, but it appears the way these models write/talk is rather deeply ingrained, and it usually end up easier for me just to write myself. Has anyone found a good approach to make models write better?

[devilsdata]

I'm a fiction writer. The way you make AI better at writing prose is the same way you make a fiction writer better at writing prose...

You edit it manually.

That's it. There are no shortcuts to good prose, except perhaps paying a ghost writer.

[jimnotgym]

It is so rare that humans write good prose, how could you expect AI to do it? It has been trained on humans of all levels

Good human prose is rarely written in one shot. It is heavily rewritten over a long period. How about getting one ai to write it, one to improve?

[iExploder]

Pompous and full of air really sums it up well. As such I usually have to completely ignore and rework generated README / code comments.

Is this just an elaborate scheme to increase in/out token churn or are we hitting some kind of LLM limitations I wonder?

[hankbond]

What is your goal? What are you trying to have it generate that is not more appropriate for you to write yourself?

[bjourne]

Hard to say since you are not presenting any concrete examples.

[drumih]

there is a skill named "avoid ai writing" on github it won't make your text as good as human text, but it will help with the most visible red flags