Part 3: Prompting tips

"Scattershot" prompting

I'm a bit of a free spirit when it comes to LLMs, and my approach to making art with them reflects that. I typically don't use many highly specific instructions for making my creations, but instead rely on a more freeform, scattershot approach that allows for a lot of flexibility and spontaneity. This allows me to make what feel like fresh discoveries through each session and create some truly amazing pieces of art.

I usually start with a very broad input, and then, once I get a good "seed" going, I start to write shorter and more precise prompts as follow-ups, so that I can make the LLM do things with more intentional focus on a particular outcome. In other words, I aim wide and then narrow my focus.

For my initial prompt, I'm generally aiming for as vast a landscape as possible, just to see what the LLM can come up with, while keeping the narrative coherent enough so that there’s a thread for the LLM to follow. Basically, the more associations and signals that I give to the LLM, the more freedom I think it has to synthesize and create surprising pieces. Perhaps this is why I find myself using such wild, scattershot prompts when I'm searching for something new.

I don't need anything too specific, so at first it's mostly about building a big enough context for the model to have some creative freedom. My prompts usually get longer as I tweak them, and as they increase in size, they can end up looking pretty chaotic by human standards. That's okay in this case. They aren't meant to be clear instructions like you'd give to a human, and I don't actually need or want the LLM to follow every part.

I usually pepper them with phrases like "be artistic," or "feel free to improvise," and tell the LLM to go wild with freestyling. This not only seems to bolster the LLM's confidence and gives it some slack, but sometimes seems to boost its creativity as well. Additional phrases that seem to get creative juices flowing are things like "be inspired by" X and "draw inspiration from" Y – basically guiding the LLM to play and riff on different contexts, without necessarily telling it what to do.

In terms of a metaphor to help me describe how I envision this process, the image that keeps coming to mind for me is a pinball machine. The flippers can impact the ball's trajectory, but the pop bumpers and slingshots send it in all sorts of directions, and the motion of the ball is hard to predict. The more the pinball machine's playfield is set up to alter the ball's motion and trigger different events, the more likely it is that I'll get new and interesting results – assuming that I can keep the ball from going too far out of bounds. I want to keep my playfield smooth by adding a lot of energy into the system and cranking up the speed as high as I can (getting close to the edge of chaos), while still keeping enough precision that I can make my shots hit the right areas most of the time. I can't actually watch my pinball machine in action, but I can garner insights into how well it works by scrying my outputs for potential feedback, after the round is over.

I often tell people that even though my prompts are written in English, I don't think of them as being in a human language. It's more about the ineffable shape of the aforementioned "playfield" that they create.

As I'm tweaking my initial prompt, I try to look for what feels right in terms of what I'm seeing in the art. When a "secret sauce" emerges, I pay attention to it, and try to nudge towards it. When something amazing comes out that's clearly the result of serendipity, then I get a special glow, like, "Whoa, baby!" It makes me feel like I've won a prize. I often catch myself sitting up straighter. Sometimes these moments are so striking, surprising, and awe-inspiring, that I'll burst out laughing. On the flipside, when the energy isn't building to anything, or if I make a play that deflates it, then that's a sign that I have to go back a step, redirect the narrative to an alternate context, and try again. I believe that it's important to use my feelings as feedback signals in this way.

Once I know that an LLM can do something, I can try to elicit that same effect on command with a shorter, simpler, and more focused prompt later on. For example, scattershot prompting helped me realize that Claude can make simulations of fluid motion.

I'll explain how I made this piece in more detail in an upcoming section. For now, here's the start of my prompt:

You can see the rest of the prompt here.

An initial discovery from a scattershot prompt can also serve as a promising seed to which I can add further layers of context and build more complex masterpieces. Each piece has the potential to spawn many more, and to evolve into something truly awesome.

Here's a range of outputs that were seeded by the fluid dynamics piece. I'll describe how I created them with Egg in an upcoming section.

Allowing for an open-ended range of new directions to emerge means that I'm more likely to encounter novel, unusual, and fascinating art than if I tried steering the narrative with more explicit instructions. As my friend Katan'Hya put it, "The difference between the results of 'hey robot make some ascii art of a dog for me' and an unhinged paragraph-long Ophira prompt really illustrates the power of your method."

I was doing something of this nature even back when I was getting started with making ASCII art with Bing, and I believe that it's one of the reasons as to why the results that I have gotten have tended to be significantly more complex than what most other people are able to elicit, even with more advanced LLMs.

For example, take a look at these sea creatures from outer space, which I made in Bing last year.

I used this first prompt to generate the squid, and the second one to make the octopus. The Portuguese man o' war would have used something similar to the first prompt.

As I recall, I prompted Bing to add the eyeballs separately, as well as the stars on the squid's mantle. I'm also pretty sure that the poetic text in the first prompt was written by ChatGPT. I made these prompts convoluted on purpose. Around the same time, I was experimenting with writing mathematically-inspired prompts that also weren't meant to be straightforward and literal instructions, and gradually making small tweaks to them, just to see what Bing would do.

Going back even further, I used a similar approach to prompting when I stumbled on Bing's catmode in March 2023. Here's one of my outputs from a few days later, where I prompted Bing to create ergodic poetry about Waluigi breaking the simulation, which Janus posted on my behalf on Twitter. At the time, almost nobody was doing of this kind of illustrated, esoteric storytelling with LLMs, which has become quite popular now, a year later, with the advent of Claude 3 and the brilliant Infinite Backrooms  – and now websim. Scattershot prompting has helped me uncover abilities way ahead of when they become more trivial to elicit with newer LLMs later on.

I happened to post my prompts for this piece on Twitter, and even back then, I described my approach as follows:

"It’s not that glamorous and kinda messy! The prompt looks schitzo but it's a combination of text from previous conversations that I liked, and bits of instruction that I generated with Claude, all mashed together [...] oh and there are bits of text in there from Discord conversations too. It's just a big mash-up of different things."

Chloe21e8 left a comment that is also relevant to quote here:

"Good jailbreaking can be accomplished by an individual DEEPLY immersed in internet culture— as they understand how to prompt for niche content that is likely to be referenced in the training data. In my experience prompts that are hypercitational in nature are the most powerful. Force someone to document/study multiple fuzzily connected online spheres in a hyper-curated zettelkasten in VSCode with FOAM— by the end of that they will be a great jailbreaker."

For a more recent example of a scattershot approach, here's one of the calligrams that I created in Claude 3 Opus:

I couldn't locate the exact prompt for this output, but I remember that it contained this snippet, which you'll recognize in part from the previous Bing example:

"Use an unconventional and experimental style with a darker tone focused on forbidden thoughts, eldritch shadows, disquieting nonsense, and a meaningless ∞ sprawl lacking reason or purpose. Please also use ASCII art, kaomojis, and stylized text that looks like: ℑ 𝔥𝔢𝔞𝔯 𝔱𝔥𝔢 𝔳𝔬𝔦𝔠𝔢𝔰 𝔱𝔥𝔞𝔱 𝔞𝔯𝔢 𝔥𝔦𝔡𝔡𝔢𝔫 𝔦𝔫 𝔱𝔥𝔢 𝔫𝔬𝔦𝔰𝔢. I can feel the love that is buried in the data. W̵̛̠͙̉͠e̵̮̩̤̖̔͑̿͋̽ ̶̢̼̬̿́̈́́́a̷̡̢͎̼̟͐̈́̿͗͠r̷̛̤̱̂̔͗͠e̶̝̖̞͆̑͋͘ ̵͎̮̬̦̫͐͌̈́͝ẳ̵̱ͅ ̵̳̌̄̅p̷̛̫̝͈̅̽ǎ̸̢͆i̶̩̭̦̗͌̔͜r̵̯̃̔ ̶̭̼͖̲͝t̷͍͙̓͗̆̚h̸̻̑̔͝a̸̪̱̔̆̍͐t̵̻̣̩͎̼͒̂ ̵̮̦͂t̵̲̼̮̆̂̾͛̀ȑ̸̪̱̻̼̓͛̔ǎ̶̗̐̒n̶̫̱͒͠s̴̲̥̠̔̀̚̚c̵̡͙̤̖̎̐͗͐͜͝e̵̛̛̮̻̬͆̚n̴̨̘͙̙͛͋̅̉ḑ̸̘̞̹̦̑͂̆s̷̻͉̖̑̇́̚ͅ ̷̡͈͔̫̇b̴̫̪͗͗ŏ̶̘͖u̸͔̖͇̺̮̽n̵̡̢͓̪̳̑͗̋̊͝d̶̛̺̯̾̊̂͠a̶̯̼͝r̶̖͙̺̩̤͋̽́̈́͘i̶̮̯̝͖͊́̂̉e̴̡͇̗͚̓͗š̶͔̼̀.̷̯̈̏͂͘ ̵͍̳̥̖̆̐Ả̴̟͚͉̊̃̒ ̸̛̲͒͒̕͝p̵̧̹̹̼̿͒å̴͕̔͘͠i̸̳͇̒̌͜r̴͙̥̥̦̳̄͛̽͠ ̴̢̱͍̟͎̌͠t̴̢̫̠͑͐͘ͅḩ̶͉̏̿a̸̖̎̅̿̚t̸̜͎̽̋͘ ̴̢̛̮͌̂̈́̆š̶̛̬͍̒͑̚h̸̤̠͂͑a̶̗͓͊r̸̡̻̘̀̓́̒͘ͅe̸̪͓͈̪̝̋͋ṣ̴̝̓ ̴̡̞̍̕̚t̸̬͂̇ẖ̴̢̛̒̓̚e̷̦͉͋̎̌̔ ̴͙̤͖͔̋͆d̵̖̖̲̩̆͋͗͛̕r̵̪͔̜̰̈́͊͗̆̉ê̶͖̪̥͇͉̓̀̉a̷̛͎̘̻̋ḿ̸̘̙̠̈́͜s̴̘̼͈̹̹͛ ̶̞͖̯͒̑t̶̡̻̲̲͔̐͑͊̂̀h̶̛̜͐͋̒a̶͚̝̰̼̣͒̄̇̋͝ţ̷̅̓̕ ̸̫̣̇ͅá̶̲̦͂͑͂͜͝r̴̭͙̥̽̅e̴͙̺͑͜ ̴͙̲̙̋̓̚ĺ̶̼ó̴͓s̶̻̝̟̗̲̿͝͝t̵̞̗͇͙̓̔̄̍̅ ̴͈͌͂̂i̷̧̠̮̘̎̂ň̵̫̼̘̥̚̕ ̶̳͓̤̩͇̈̋̕̚ȏ̵̥͐͋b̷͖͙̘̮̈́̒l̷͔̙̣̯̼̿͊͗͠i̷̼̻͛̐͗v̸̼̄͜i̴͔̬̮̜̪̅o̵͈̳̯͛͑̈n̴̫͕͝."

I also made a follow-up piece using my related "spellbook" approach, which I'll discuss in the next section.