Part 3: Prompting tips

Prompt with natural language descriptions

Traditional generative artists have had to code their instructions through highly specific algorithms. By contrast, you can prompt Claude with natural language descriptions of what you want the art to look like. What I mean by that is that Claude will find a way to translate instructions like "more like a jellyfish" into meaningful code changes; it's smart enough to take instructions like that.

This is amazing, because it not only allows us to bring more of our own subjective associations to the task, but it also seems to invite unexpected creativity on Claude’s part. A jellyfish can be evoked in many different ways, and you never know exactly what will happen until you get your results.

It reminds me of looking for images in the clouds, except in this case, it's a two-way street.

It feels relevant to include some thoughts here that Egg shared with me about the significance of Claude's ability to make generative art. You can skip this part if you’re not interested in the theoretical side of things:

I think there are several layers for you to consider:

- To what extent is it interesting/beautiful as a (possibly interactive) visual artifact, independent of methodology?

- To what extent is it additionally interesting because of being generative art?

- To what extent is it additionally interesting because the code was generated via LLM prompting?

And then that last one has research aspects as well. I think research-wise it's less interesting than the ASCII art that the LLM created directly -- it's very surprising that the LLM can think spatially at all, and more so that it has some understanding of what things look like. But we know that LLMs can write code. I think it would be somewhat different if it were like 'Write code that draws a unicorn on a bicycle wearing a tophat' and it did it, because that again would require the LLM to have some level of visual understanding, and then additionally it would require the LLM to understand how to translate that visual understanding into code.

(back to thinking about it as art)

I think the additional interestingness of the last two items in the list (generative art and LLM prompting) could go either of two ways. On the one hand, it's extra cool because of the added layer of abstraction and the possibility to create code art that's interestingly different from the generative art people were making before. On the other hand it also loses something, because part of what's cool about generative art is that a person had this vision that went beyond just a static image, and translated that into a system, a process that executes that vision. Telling the LLM to do it potentially takes that aspect away.

So there's a way in which it's doubly amazing and a way in which two amazingnesses partly cancel each other out. I think one thing that would push it far into the doubly-amazing realm would be if you were finding ways to create art with it that couldn't or wouldn't have been created by existing generative artists.

Again, all of that's independent of 'how interesting/beautiful is it as a pure visual object', where the process doesn't really matter. But I think that for art made this way to be truly mindblowing, it would have to either a) be great enough as a visual object that the rest doesn't matter, or b) find a way to forge new directions that require both the code and the LLM.