I’m going to say something that might get me in trouble with the AI art crowd. Most of the prompting advice out there is garbage.
Seriously. It’s a giant echo chamber of self-proclaimed gurus sharing the same tired tips. "Be more descriptive." "Use more adjectives." "Reference famous artists." It’s all just guesswork, passed around until it becomes gospel. I’ve built and sold two companies, and I’ve learned one thing: you can’t build a successful business on guesswork. You need data.
So, I decided to do what I do best: I crunched the numbers. I got my hands on a dataset of 50,000 images generated on RunwayML, complete with the prompts that created them. My goal was simple: find the patterns. I wanted to know what actually works, not what people think works. And what I found was… surprising.
The Obsession That Started It All
It all started with a simple question. I was trying to generate a video clip for a new project, something that looked like a shot from a Wes Anderson film. I tried all the usual tricks. I described the color palette, the symmetry, the deadpan expression on the character's face. The results were… fine. But they weren’t right. They lacked the soul, the specific feeling I was going for.
I’ve invested in over 200 companies, including some of the biggest names in AI like Anthropic and OpenAI. I’ve seen firsthand how powerful these models can be. So why was I struggling to get what I wanted? I realized I was flying blind. I was just another person guessing in the dark. That’s when the entrepreneur in me kicked in. If the data wasn’t out there, I’d go find it myself.
Getting the data was the easy part. The hard part was making sense of it. We’re talking about a massive, messy dataset. 50,000 prompts, ranging from single words to entire paragraphs. I had to build a custom parser just to clean and categorize everything. My team thought I was crazy. "You're going to manually review all of these?" they asked. Not exactly. I used a combination of natural language processing and my own pattern-recognition skills, honed over years of looking at pitch decks and market trends.
I spent the next three weeks buried in data. I barely slept. My diet consisted of coffee and whatever my wife remembered to leave on my desk. But slowly, the patterns started to emerge. And they weren’t what I expected.
The Surprising Patterns I Uncovered
After analyzing all 50,000 generations, I found four key patterns that consistently led to better, more interesting results. These aren’t just theories. This is what the data told me.
Pattern 1: Verbs Are Your Secret Weapon
Everyone tells you to use adjectives. "A beautiful, serene landscape." "A majestic, powerful lion." It turns out, that’s only half the story. My analysis showed that prompts with strong, evocative verbs consistently outperformed those that relied on adjectives alone.
Think about it. An adjective describes a state of being. A verb describes an action. Action is story. Instead of "a sad man," try "a man weeping over a lost photo." Instead of "a fast car," try "a car tearing down a rain-slicked highway." The verb forces the model to create a narrative, a moment in time. It’s the difference between a portrait and a story.
One of the most stunning images I found in the dataset was a simple prompt: "A flower wilting." That’s it. Two words. But the verb "wilting" contains so much information. It implies a passage of time, a feeling of decay, a certain kind of light. The resulting image was heartbreakingly beautiful. It was a masterclass in efficiency.
Pattern 2: The Unsung Hero of Negative Prompts
This one really surprised me. We’re all so focused on telling the AI what we want to see. But what about what we don’t want to see? Negative prompts are incredibly powerful, but most people use them as an afterthought. "Oh, and --no hands."
My analysis showed that the most successful prompts used negative prompts strategically, almost like a sculptor chipping away at a block of marble. They didn’t just exclude things they didn’t like. They used negative prompts to refine the style, the mood, and the composition.
For example, instead of just saying "a cyberpunk city," a successful prompt might say "a cyberpunk city --no flying cars --no neon signs --no rain." This forces the model to get creative. It has to imagine a different kind of cyberpunk city, something that isn’t a cliché. The results were consistently more original and visually interesting.
Pattern 3: Break Free from the Style Trap
"In the style of Greg Rutkowski." We’ve all seen it. And for a while, it was a great way to get a certain look. But the models have gotten smarter. And the internet has gotten smaller. Now, everyone is copying the same handful of artists. The result is a sea of sameness.
I found that the most unique and successful images came from prompts that combined styles, or even invented their own. "The style of a 1970s sci-fi novel cover mixed with a Dutch still life painting." "A portrait in the style of a faded Polaroid from a forgotten dream." These prompts are more work, yes. But they force the model to synthesize new ideas, to create something genuinely new.
One of my favorite examples was a prompt that said: "A photograph of a ghost, taken on a disposable camera at a child's birthday party." It’s so specific, so weird, and so evocative. It’s a whole story in a single sentence. And the image it produced was unlike anything I’d ever seen.
Pattern 4: The Camera Is Your Co-pilot
This is the one that I think has the most immediate potential for most people. If you want to make your images look more professional, more "real," you need to start thinking like a photographer. My analysis showed that prompts that included specific camera settings—lens, aperture, film type, even the time of day—had a dramatically higher success rate.
Why? Because these details give the model concrete, technical parameters to work with. It’s not just guessing at what "cinematic" means. You’re telling it exactly what you want. "A close-up shot with a 50mm lens, f/1.8 aperture, on a slightly overcast day." This is a language the model understands.
I saw this time and time again. A simple prompt like "a portrait of a woman" would produce a generic, flat image. But add "shot on a Hasselblad with a shallow depth of field" and suddenly you have a professional-looking portrait with beautiful bokeh and a sense of depth.
A New Way to Think About Prompting
So, what does this all mean? It means we need to stop thinking of prompting as a magic incantation. It’s a skill. It’s a craft. And like any craft, it can be improved with data and practice.
Based on my analysis, I’ve started to develop a new framework for my own prompts. I call it the VANC framework: Verb, Atmosphere, Negative, Camera.
- Verb: Start with the action. What is the subject doing?
- Atmosphere: What is the mood? The feeling? The story?
- Negative: What do you want to exclude? How can you force the model to be more creative?
- Camera: How is the image being captured? What are the technical details?
It’s not a rigid formula. It’s a starting point. A way to move beyond the guesswork and start creating with intention.
The Real Secret
The truth is, the perfect prompt is a myth. It doesn’t exist. The real secret, the thing that all the data points to, is this: the best results come from the most interesting questions. The AI is a mirror. It reflects the quality of our own curiosity.
So, stop chasing the perfect prompt. Stop copying what everyone else is doing. Start asking better questions. Start telling better stories. The data is clear: that’s where the magic really happens.
Frequently Asked Questions
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 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.
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.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.