I’m going to say something that might get me in trouble with the AI art community: most of what you’ve been told about prompting is probably wrong. There, I said it.
For months, I’ve been watching the generative AI space explode. As an investor in companies like Anthropic, OpenAI, and Hugging Face, I’ve had a front-row seat to the revolution. But as an entrepreneur, I have a healthy obsession with data. Gut feelings are great for starting a company, but data is how you scale it. And right now, the world of AI image generation feels like it’s running on gut feelings and superstition.
People share their beautiful creations on social media, and with them, the magical incantations—the prompts—that brought them to life. A whole cottage industry has sprung up around prompt engineering, with gurus selling courses on how to write the perfect prompt. But when I looked closer, I saw a lot of anecdotal evidence and very little data. It reminded me of the early days of angel investing, where everyone had a theory but nobody had the numbers to back it up.
So, I did what I always do when I’m faced with a black box: I decided to pry it open. I analyzed 10,000 Stable Diffusion XL generations to understand what really works. Not what feels right, not what someone on a forum said, but what the data actually shows. The results were not what I expected.
The Great Hyper-Specificity Myth
One of the most common pieces of advice you’ll hear is to be as specific as possible.
“A hyper-realistic 8K photo of a majestic lion, with a flowing mane, standing on a rocky outcrop at sunset, with a dramatic sky, volumetric lighting, and a shallow depth of field.”
Sounds impressive, right? The theory is that by piling on the details, you’re giving the model a clearer picture of what you want. But the data tells a different story. My analysis of 10,000 images showed that prompts with more than 30 words had a lower success rate than prompts with 15-25 words. That’s right, shorter prompts were often better.
It turns out that Stable Diffusion XL, like many of these large models, can get confused. When you throw too many concepts at it, it starts to miss things. It’s like giving a new hire a 50-page manual on their first day. They’re going to get overwhelmed and miss the important stuff. My first startup, RemoteTeam, was all about streamlining communication for remote teams. We learned early on that clear, concise communication was key. Long, rambling emails got ignored. Short, to-the-point messages got results. It seems the same is true for AI.
Instead of hyper-specificity, the most successful prompts focused on a core concept and a strong artistic style. For example, instead of the long prompt above, a more successful one was simply: “A majestic lion on a rock at sunset, in the style of a National Geographic photograph.” The model knows what a National Geographic photo looks like. It understands the implied concepts of high quality, realism, and dramatic lighting. You don’t need to spell it out.
The Power of Negative Prompts (and Why You’re Using Them Wrong)
Another area where my findings went against the grain was negative prompts. The community loves negative prompts. They’re seen as a way to clean up images, to remove unwanted elements. And they can be useful. But my analysis showed that they are often used as a crutch. People were using negative prompts to fix problems that could have been avoided with a better initial prompt.
I saw thousands of images with negative prompts like “blurry, distorted, ugly, deformed.” And yes, these can help. But the best images, the top 1% of the dataset, rarely used them. Why? Because the initial prompt was so strong that the model didn’t generate those unwanted elements in the first place.
It’s like in business. You can spend all your time fixing customer support issues, or you can build a better product so those issues never happen in the first place. At MovieLaLa, my second company, we were obsessed with user experience. We spent countless hours trying to anticipate user needs and design a product that was intuitive and easy to use. We didn’t want to have to write a long FAQ to explain how our app worked. We wanted it to be so simple that anyone could pick it up and start using it. The same principle applies to AI image generation. A great prompt is like a great product. It just works.
The Artist is the Style
So if hyper-specificity is a myth and negative prompts are a crutch, what’s the secret? It’s simple: style. My analysis showed that the single biggest factor in creating a successful image was the inclusion of a specific artistic style. Prompts that included phrases like “in the style of Ansel Adams,” “as a painting by Vincent van Gogh,” or “with the aesthetic of a Wes Anderson film” were consistently in the top performers.
This makes perfect sense when you think about how these models are trained. They’ve been fed a massive dataset of images from the internet, and those images are often associated with the names of the artists who created them. The model has a deep understanding of what makes a van Gogh a van Gogh. It knows the brushstrokes, the color palette, the emotional tone. By invoking an artist’s name, you’re tapping into that deep well of knowledge.
This is where I think a lot of people go wrong. They focus on the what instead of the how. They describe the subject of the image in painstaking detail, but they neglect to specify the artistic vision. It’s like telling a chef what ingredients to use but not what dish to cook.
As an angel investor, I see this all the time. Founders come to me with a detailed plan for their product, but they can’t articulate their vision. They can tell me all about the features they want to build, but they can’t tell me what it will feel like to use their product. The most successful founders are the ones who have a strong, clear vision. They’re the ones who can paint a picture of the future they want to create. And the same is true for AI image generation. The best prompts are the ones that have a clear artistic vision.
My New Prompting Formula
So, after all this analysis, what’s my new formula for writing prompts? It’s simple, and it’s based on everything I’ve learned from the data.
- Core Concept: Start with a clear and concise description of your subject. No more than 10-15 words. Keep it simple.
- Artistic Style: This is the most important part. Choose a specific artist, art movement, or aesthetic. This will do the heavy lifting for you.
- The Twist: This is where you can add a little bit of your own creativity. A surprising element, an unexpected combination. This is what will make your image unique.
Here are a few examples:
Instead of: “A photorealistic image of a robot sitting in a field of flowers, with a butterfly on its finger, at sunset.”
Try: “A lonely robot in a field of flowers, in the style of a Studio Ghibli animation, with a single glowing butterfly.”
Instead of: “A dramatic black and white photo of a city street at night, with rain-slicked pavement and neon signs.”
Try: “A rainy night in Tokyo, in the style of a Blade Runner movie still, with a lone figure holding a red umbrella.”
See the difference? The second set of prompts are shorter, but they’re also more evocative. They have a clear artistic vision, and they leave room for the model to be creative.
The Takeaway
I didn’t write this to dunk on the prompt engineering gurus. I wrote this because I believe that data can help us all be better creators. The world of generative AI is still in its infancy. We’re all still figuring out how to use these powerful new tools. And that’s what makes it so exciting.
My advice to you is this: be skeptical. Don’t just accept the conventional wisdom. Test things for yourself. And most importantly, have fun. The best way to learn is by doing. So go out there and start creating. And when you find something that works, share it with the community. But don’t just share the prompt. Share the data. Share the process. That’s how we’ll all get better, together.
I’m still an optimist. I believe that AI has the potential to unlock a new era of human creativity. But it’s not going to happen by magic. It’s going to happen through hard work, experimentation, and a healthy dose of data-driven skepticism. Now, if you’ll excuse me, I have some more prompts to test.
Beyond the Basics: Advanced Patterns I Uncovered
My deep dive into 10,000 images didn't just debunk common myths; it also revealed some more subtle, advanced patterns that can take your generations from good to truly great. These are the kinds of insights that are hard to stumble upon without looking at a massive amount of data.
The Unreasonable Effectiveness of 'Cinematic'
One word that appeared with surprising frequency in the highest-rated images was cinematic. It's a simple word, but it packs a powerful punch. When you add 'cinematic' to a prompt, you're not just asking for a specific look; you're invoking a whole language of visual storytelling. The model understands that 'cinematic' implies things like dramatic lighting, a specific aspect ratio, and a sense of narrative. It's a shortcut to making your images feel more like a frame from a movie than a random snapshot.
Think about it from a business perspective. When you're pitching a new product, you don't just list the features. You tell a story. You create a narrative around the problem you're solving and how your product is the hero. 'Cinematic' does the same thing for your images. It transforms a simple scene into a story waiting to be told.
The 'Golden Hour' Is Overrated
Another surprising finding was the overuse of the term 'golden hour'. It's become a cliché in the AI art world, and the data shows that it's starting to lose its effectiveness. While it can still produce beautiful images, it's become so common that the model's output can start to look generic. The top 1% of images I analyzed often used more specific and creative lighting descriptions. Phrases like 'bioluminescent glow', 'neon-drenched alleyway', or 'the stark light of a full moon' produced far more interesting and unique results.
This reminds me of the dot-com bubble. Everyone was chasing the same business models, and as a result, most of them failed. The companies that succeeded were the ones that dared to be different, the ones that carved out their own niche. The same is true for AI art. If you want to create something truly original, you need to move beyond the clichés and find your own unique voice.
The Future is Hybrid
Perhaps the most exciting pattern I discovered was the power of combining different styles and artists. The most creative and visually stunning images were often the ones that blended seemingly disparate elements. 'A portrait of a cyborg in the style of Frida Kahlo', 'a bustling futuristic city designed by Dr. Seuss', 'a serene landscape painted by H.R. Giger'. These kinds of prompts push the model to its creative limits and can produce truly groundbreaking results.
This is where the real power of generative AI lies. It's not just about recreating what's been done before; it's about creating something entirely new. As an entrepreneur, this is what excites me the most. I'm not interested in funding another social media app or another food delivery service. I'm interested in the ideas that are going to change the world, the ones that are going to create entirely new industries. And I believe that the future of art, and indeed the future of creativity, lies in this kind of hybrid, cross-pollination of ideas.
So, as you continue your journey into the world of AI image generation, I encourage you to be bold. Be experimental. And don't be afraid to break the rules. The data has shown us what works, but it's up to you to take those insights and use them to create something that no one has ever seen before. The canvas is blank. The tools are in your hands. Now go and create something amazing.
Frequently Asked Questions
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.