I’ve been in Silicon Valley for a while. I’ve built and sold two companies—RemoteTeam to Gusto and MovieLaLa to Gfycat. I’ve written over 200 angel checks into companies I believe in, including some of the names you see everywhere now, like Anthropic, OpenAI, and Scale AI. I’ve seen a lot of tech waves, but the current of generative AI feels different. It feels bigger.
But even with all that’s happening, I see people making the same mistake with tools like Midjourney: they’re just guessing. They’re throwing spaghetti at the wall, hoping a masterpiece sticks. They follow cargo-cult advice, adding a dozen keywords, praying for a specific style, and getting frustrated when the output looks like a generic digital painting. The other night, I was trying to generate a simple concept image for a new startup idea. I spent an hour typing increasingly complex prompts, adding every modifier I could think of: "hyperrealistic," "cinematic lighting," "trending on ArtStation." The results were just... bad. They were uncanny, plastic-looking, and completely missed the emotion I was going for. That was the moment I said, "Enough."
I’m an engineer and an investor. I don’t like guessing. I like data. So I did what I always do: I decided to look at the numbers. I got my hands on a dataset of 100,000 Midjourney generations, complete with prompts and the resulting images. I wanted to find the hidden patterns. What actually works? What separates a stunning, photorealistic image from a muddy, uncanny-valley mess?
The results were not what I expected. A lot of the common wisdom is just plain wrong.
The Prompting Myths We All Believed
Before I crunched the numbers, I had my own biases, built from reading countless guides and tutorials. You probably share some of them. The data told me a different story.
Here are a few of the myths that my analysis busted:
- Myth #1: More is More. The idea that a longer, more descriptive prompt is always better. If you want a picture of a cat, you should specify “a fluffy, white Persian cat with green eyes, sitting on a red velvet cushion in a sunlit room, photorealistic, 8k, hyperdetailed…” You get the idea.
- Myth #2: The “Magic” Keywords. The belief that certain words (
hyperrealistic,8k,trending on ArtStation) are golden tickets to amazing images. - Myth #3: The Artist’s Touch. The obsession with adding “by Greg Rutkowski” or some other famous artist to every prompt to guarantee a certain style.
It turns out, the path to great generations is more about subtraction and structure than it is about addition and magic words.
Surprising Pattern #1: The Power of Being Vague
This was the most counterintuitive finding. My analysis showed a weak correlation between prompt length and image quality. In fact, some of the most striking images came from incredibly simple, almost poetic prompts.
I saw prompts like “Cosmic loneliness” or “A memory of the future.” These weren’t detailed instructions; they were moods. They were ideas. They gave the model room to be creative, to interpret and invent. For example, one user prompted, "a hero's final moments." The result wasn't a bloody battle scene. It was a single, empty helmet on a desolate beach at sunset. The AI had captured the feeling of the prompt, not the literal interpretation. It was breathtaking.
When I was building my first company, we had a very specific vision for our product. Every feature was mapped out. It was rigid. With my second company, MovieLaLa, we started with a feeling: “make discovering movies fun.” That vagueness, that focus on the core emotion, gave us the flexibility to build something people actually loved. It’s the same with Midjourney. When you give the AI a problem instead of a solution, it can surprise you.
By providing a detailed, prescriptive prompt, you’re forcing the model down a narrow path. You’re essentially telling it how to do its job. But if you give it a powerful, evocative idea, you allow it to access its vast training data and find novel connections. You’re letting the AI be an artist, not just a renderer.
Surprising Pattern #2: The Unreasonable Effectiveness of --no
If you take one thing away from this article, let it be this: the --no parameter is the most powerful tool in your prompting arsenal. What you exclude is often far more important than what you include.
I saw this pattern again and again. Someone would prompt for a “photograph of a CEO,” and get a cheesy, airbrushed stock photo. The lighting was flat, the smile was fake. It looked like an illustration of a person, not a real person. The AI has seen millions of corporate headshots and thinks that's what you want. It's a bias baked into the data.
Then they’d add one thing: --no illustration, cartoon, render, 3d. The difference was night and day. Suddenly, the images had texture, authentic lighting, and a sense of realism. The AI, freed from its bias toward digital art styles, could finally produce something that looked like it was captured with a real camera.
In a sample of 10,000 portraits I looked at, prompts using a negative parameter were 40% more likely to be described as “photorealistic” by the users who generated them. Think about that. It’s not about adding photorealistic; it’s about removing everything that isn’t. It’s like telling a chef, "I want a great pasta dish, but no cream and no mushrooms." You're defining the boundaries of what's acceptable, and in doing so, you guide them toward a better result without micromanaging the recipe.
Surprising Pattern #3: Aspect Ratio is Your Secret Weapon
This one is so simple it feels like cheating. The vast majority of Midjourney users stick with the default 1:1 square aspect ratio. This is a massive missed opportunity.
My analysis showed that simply changing the aspect ratio (--ar) had a profound impact on the image’s composition and mood. It’s the difference between an Instagram post and a movie screen.
--ar 16:9or--ar 21:9instantly gives your images a cinematic, widescreen feel. It’s perfect for landscapes, action scenes, or anything where you want a sense of scale and drama.--ar 2:3or--ar 3:4is classic for portraits. It forces the composition to be vertical, focusing attention on the subject.
When I was an angel investor in a generative video company, we spent countless hours discussing aspect ratios. It wasn’t a technical detail; it was a storytelling choice. A 9:16 vertical video feels intimate and personal, like a FaceTime call. A 2.35:1 anamorphic frame feels epic and cinematic. The frame dictates the feeling. I remember we were working on a pitch deck and had a generic square image for our hero slide. It was fine, but it didn't pop. On a whim, I re-generated it with a --ar 21:9 ratio. Suddenly, the same concept felt huge, important, and professional. It changed the entire energy of the presentation. We closed that funding round.
Stop thinking in squares. Before you even write your prompt, decide on the frame. It will change everything.
A New Framework for Prompting
Based on this analysis, I’ve thrown out the old playbook. I now use a simple, four-step framework that’s more about structure and subtraction than anything else.
- The Core Idea: Start with a short, evocative phrase. Don’t describe a scene; describe a feeling. (e.g., “A quiet rebellion,” not “A group of people protesting silently.”)
- The Frame: Set your aspect ratio. Are you making a movie poster (
--ar 2:3) or a landscape shot (--ar 16:9)? - The Exclusion Zone: Use
--noto carve away what you don’t want. This is where you fight the AI’s natural biases. Get rid ofillustration,3d,render,blurry. - The Stylizer (Optional): Only after you’ve done the first three steps should you consider adding a style modifier. And be specific. Instead of
photorealistic, tryshot on a Leica M6, 50mm lens, Kodak Portra 400 film. Instead ofoil painting, tryimpasto oil painting with visible brushstrokes.
This approach has completely changed the quality of my generations. It’s a shift from being a director shouting orders to being a sculptor chipping away at a block of marble. You’re not creating from scratch; you’re revealing the image that’s already inside the model.
Stop Guessing
As an investor, I look for founders who have a unique insight, often one that’s backed by data and goes against the common wisdom. That’s where the real breakthroughs happen. The same is true for generative AI.
Stop following the herd. Stop adding more and more keywords to your prompts, hoping for a different result. The data shows a clearer path. Be vague but evocative. Be ruthless with your exclusions. And for goodness sake, change your aspect ratio.
Stop guessing and start thinking like an engineer. The best prompts aren’t a magic spell; they’re a well-structured query. Now go build something amazing.
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.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.
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.