I've been a fan of 'random text' for some time. "Random text" is a bit vague, but to me the idea of using code to generate random stories, or even snippets, is fascinating. Back in April, I blogged about how I created short dragon-based stories. It took a generic string:

A #adjective# dragon lives #place#. She #verb# her hoard, which consists of a #number# of #thing#, #number# of #thing#, and #number# of #thing#. She feels #feeling#.

And created a story by replacing the pound-wrapped tokens with real words. I used a couple of different tools to build this, but the core one was a cool little Node library named random-word-slugs. It's a powerful random word library that can get pretty specific in terms of focusing on a particular type of word. For example, here's the code I use to generate an adjective for the dragon focused on three aspects - color, appearance, and personality:

const getDragonType = function() {
  const options = {
    format:'lower',
    partsOfSpeech: ['adjective'],
    categories: {
      adjective: ['color','appearance','personality']
    }
  }
  
  return randomWordSlugs.generateSlug(1, options);
}

I was curious if I could take this approach and build it with generative AI. Here's what I found.

Version One

For my first version, I opened up Google AI Studio, started a new prompt, and used these system instructions:

You are a text parser. Given a prompt with variables wrapped in {{ and }}, you will replace that with a value based on the contents of the text inside the brackets. So for example, {{noun}}, means to replace it with a random noun, and {{animal}}, means to replace it with a random animal.

With this in place, I then did a test:

This is a story about a {{noun}}. It was colored {{color}} and liked to eat {{food}}. On the weekend, it liked to do {{activity}}.

Which gave me:

This is a story about a tree. It was colored blue and liked to eat pizza. On the weekend, it liked to do skydiving.

Not bad. Then I decided to push it a bit:

I like to eat {{food}}. But sometimes when I eat it, I feel {{sad feeling}}. My favorite holiday is {{holiday}}. My favorite song is {{song}}.

I was curious how well it would handle sad feeling and holiday. I got this:

I like to eat ice cream. But sometimes when I eat it, I feel guilty. My favorite holiday is Christmas. My favorite song is Bohemian Rhapsody.

This was kind of shocking how close it was to my own personal answers. ;) I changed sad feeling to happy feeling and got this:

I like to eat pizza. But sometimes when I eat it, I feel happy. My favorite holiday is Thanksgiving. My favorite song is Imagine.

Finally, I decided to really push it:

I like to eat {{food}}. But sometimes when I eat it, I feel {{weird feeling}}. My favorite holiday is {{holiday}}. My favorite song is {{song by The Cure}}.

It had no problem with specific song token:

I like to eat sushi. But sometimes when I eat it, I feel disoriented. My favorite holiday is Halloween. My favorite song is Friday I'm In Love by The Cure.

Pretty impressive. But then I decided to push more...

Version Two

In the past, I worked with a library that could generate a random word and then remember it for reuse later. Tracery had a feature where it could select a random word, but then remember the chosen word for use again later. I decided to see if I could get that working with Gemini. I updated the system instructions like so:

You are a text parser. Given a prompt with variables wrapped in {{ and }}, you will replace that with a value based on the contents of the text inside the brackets. So for example, {{noun}}, means to replace it with a random noun, and {{animal}}, means to replace it with a random animal. You also support storing and remembering values. If the string inside the brackets contains a colon, the value after the colon is the name of a variable. So for example, {{noun:itemX}} means to select a random noun and insert it into the result, but also store the value in a variable named itemX. If I use {{itemX}} again, you will use the previous value.

And again, it worked really dang well. I started with:

My first name is {{first name:identity}}. Say hi to me, my name is {{identity}}.

And got:

My first name is Sarah. Say hi to me, my name is Sarah.

I then tried:

My first name is {{first name:fname}}. My last name is {{last name:lname}}. I write my formally as {{lname}}, {{fname}}. My favorite color is {{color}}.

I tried this twice and got:

My first name is David. My last name is Smith. I write my formally as Smith, David. My favorite color is blue.

My first name is Emily. My last name is Jones. I write my formally as Jones, Emily. My favorite color is green.

As a last test, I tried the Dragon example above, replacing pounds with double brackets, and it did an admiral job:

A fiery dragon lives in a cave beneath a mountain. She guards her hoard, which consists of a thousand of gold coins, five hundred of precious gems, and a hundred of magic artifacts. She feels content.

I see two issues here. One, I should remove of. Second, I'd be willing to bet the place token will consistently be things relevant to dragons. I did a few more tests and it definitely seemed that way:

A sleepy dragon lives in a cozy cave. She snuggles her hoard, which consists of a thousand feather pillows, five hundred soft blankets, and a hundred stuffed toys. She feels comfortable.

So on a whim, I tried changing {{place}} to {{any place, not just one a dragon would like}}. And wow, I got some impressive results. Here's one:

A sparkly dragon lives in a crowded library. She organizes her hoard, which consists of a thousand books about dragons, five hundred maps of forgotten kingdoms, and a hundred ancient scrolls. She feels knowledgeable.

What impresses me is that the items seem to make sense for the place. I guess that's expected, but I wouldn't have been able to do that with the Node library I used.

Can You Use This?

Normally when I blog about Gemini and GenAI stuff, I have a code demo to go along with what I demonstrate in the website, but honestly, I don't think it would add much. I have to say, barring issues with the price of the API call, I think I would absolutely use a solution like this if I needed to generate a random string according to a template. I suppose it makes sense, this is generative AI after all, but I didn't expect it to work so well. Let me know what you think.