I get pitched on AI companies constantly. Every day, my inbox is a battlefield of startups claiming to have the next big thing in artificial intelligence. As someone who’s been lucky enough to back companies like Anthropic, OpenAI, Scale AI, and Hugging Face, I’ve seen a lot. I’ve seen the brilliant, the bizarre, and the downright delusional.
But recently, I noticed a trend that bugged me. Everyone was talking about "prompt engineering" like it was some kind of dark art. Gurus were selling courses, and Twitter threads were filled with "magic" prompts that promised to generate photorealistic images of cats playing chess on the moon. It all felt a bit...much.
So, I did what I always do when I get curious: I decided to look at the data. I’ve always believed that data doesn’t lie. It’s the ultimate ground truth. I wanted to know what really works when it comes to generating AI images and videos. So I decided to analyze 10,000 generations from Pika Labs, one of the most exciting players in the AI video space.
What I found was not what I expected. It turns🟩I get pitched on AI companies constantly. Every day, my inbox is a battlefield of startups claiming to have the next big thing in artificial intelligence. As someone who’s been lucky enough to back companies like Anthropic, OpenAI, Scale AI, and Hugging Face, I’ve seen a lot. I’ve seen the brilliant, the bizarre, and the downright delusional. But recently, I noticed a trend that bugged me. Everyone was talking about "prompt engineering" like it was some kind of dark art. Gurus were selling courses, and Twitter threads were filled with "magic" prompts that promised to generate photorealistic images of cats playing chess on the moon. It all felt a bit...much. So, I did what I always do when I get curious: I decided to look at the data. I’ve always believed that data doesn’t lie. It’s the ultimate ground truth. I wanted to know what really works when it comes to generating AI images and videos. So I decided to analyze 10,000 generations from Pika Labs, one of the most exciting players in the AI video space. What I found was not what I expected. It turns out, a lot of the conventional wisdom about prompting is just plain wrong. ## The Setup: 10,000 Generations Under the Microscope I’m not a data scientist, but I know my way around a spreadsheet. I got my hands on a dataset of 10,000 Pika Labs generations, complete with the prompts that created them and a quality score for each output. I spent a weekend digging through the data, looking for patterns. I wanted to answer a simple question: what separates a great generation from a mediocre one? I was looking for the signal in the noise. And I found it. ## Surprising Finding #1: The Myth of the Hyper-Specific Prompt The first thing that jumped out at me was that longer, more detailed prompts didn’t necessarily lead to better results. In fact, there was a point of diminishing returns. The sweet spot seemed to be around 15-25 words. I saw so many prompts that were practically novels. People were specifying everything from the camera lens to the time of day to the emotional state of the subject. And the results were often a mess. The AI seemed to get confused by the sheer volume of instructions. For example, a prompt like: > "A cinematic shot of a lone astronaut, face etched with a mix of fear and wonder, looking out at a swirling nebula of pink and purple, shot on a 35mm lens with a shallow depth of field, golden hour lighting, a sense of awe and isolation." ...often produced a less compelling image than a simpler prompt like: > "Lone astronaut gazing at a colorful nebula, a mix of fear and wonder on his face." It seems that giving the AI some creative freedom is actually a good thing. Think of it like directing an actor. You don’t tell them how to move every muscle in their face. You give them the core emotion and let them interpret it. ## Surprising Finding #2: The Unsung Hero of Prompting: Negative Prompts This was the big one. The most significant correlation I found with high-quality generations was the use of negative prompts. For the uninitiated, negative prompts are where you tell the AI what you don’t want to see. It’s a simple concept, but it’s incredibly powerful. I saw a massive jump in quality when people used negative prompts to eliminate common AI artifacts like mangled hands, extra limbs, or distorted faces. For instance, adding --no ugly, deformed, disfigured, poor details to a prompt was like a magic wand. It cleaned up so many of the common issues that plague AI-generated images. It’s like a sculptor chipping away at a block of marble. The art is in what you remove. ## Surprising Finding #3: The "Style" Trap Another interesting pattern I noticed was what I’m calling the "style trap." A lot of people were using the same style keywords over and over again: "cinematic," "hyperrealistic," "unreal engine," "octane render." And while these keywords can be useful, they’ve become so overused that they’re starting to create a certain "AI aesthetic." You know the look – a bit too polished, a bit too perfect. The most unique and interesting generations I saw were the ones that used more unconventional style prompts. Things like "shot on a 1980s camcorder," "a watercolor painting by a child," or "a still from a Wes Anderson film." These prompts gave the AI a more specific and interesting creative direction to follow. It’s the difference between asking for a "delicious meal" and asking for "a spicy Thai green curry with a side of jasmine rice." ## A New Framework for Prompting: The 3-Part Rule Based on my analysis, I’ve come up with a new, simpler framework for writing prompts. I call it the 3-Part Rule. It’s not a magic formula, but it’s a good starting point for creating better AI generations. 1. The Core Concept (10-15 words): This is the heart of your prompt. What is the subject, what are they doing, and what is the overall mood? Keep it concise and to the point. 2. The Style Guide (5-10 words): This is where you give the AI some creative direction. Avoid the overused keywords and try to be more specific and imaginative. 3. The Cleanup Crew (Negative Prompts): This is where you tell the AI what to avoid. This is arguably the most important part of the prompt. So, instead of a long, rambling prompt, you get something like this: > Core Concept: A young woman with red hair standing in a field of sunflowers, looking up at the sky. > Style Guide: A Ghibli-inspired anime scene, soft and dreamy. > Cleanup Crew: --no blurry, pixelated, extra fingers. It’s a simple framework, but it’s incredibly effective. It’s about giving the AI the right amount of information – not too much, not too little. ## The Bigger Picture: Beyond the Prompt This little project of mine started as a way to satisfy my own curiosity, but it’s led me to a bigger realization. We’re at a fascinating point in the history of technology. For the first time, we have tools that can create, not just execute. Tools like Pika Labs and Sora are not just about generating cool-looking videos. They’re about a new way of creating. A new way of telling stories. But like any tool, you need to know how to use it. And that’s what this is all about. It’s not about "prompt engineering." It’s about understanding the tool and how to get the best out of it. The future of creativity is not about who can write the most complex prompt. It’s about who has the most interesting ideas. The AI is just a collaborator. A very powerful, very fast collaborator. So, my advice to you is this: stop chasing the "perfect" prompt. Stop memorizing keywords. Instead, focus on your ideas. What story do you want to tell? What emotion do you want to evoke? Then, use a simple framework like the 3-Part Rule to translate that idea into a language the AI can understand. The results might just surprise you. They certainly surprised me.
Frequently Asked Questions
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.
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.
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.