I have a confession to make. I’m obsessed. When I’m not meeting with founders, reviewing pitch decks, or working on my own projects, I’m down the rabbit hole of generative AI. It started as a curiosity, a way to visualize concepts for startups I was advising. Now, it’s a full-blown addiction. My latest fixation? Pika Labs. And I went deep.
I didn’t just generate a few dozen images. I analyzed over 50,000 public generations from the platform. I wanted to see past the generic advice of “be more descriptive.” I knew there had to be underlying patterns, subtle tricks that separate the stunning, viral-worthy clips from the muddy, uncanny-valley failures. And I found them.
Most of what I discovered goes against the common wisdom you read on Reddit or in Discord channels. It’s less about cramming every possible detail into your prompt and more about psychology, suggestion, and knowing how to speak the model’s hidden language.
The Day a Bad Prompt Almost Killed a Good Idea
A few months ago, I was helping a portfolio company prepare for a crucial seed round presentation. They’re building a platform for sustainable fashion, and they wanted a powerful, evocative video for their opening. The idea was to show a dress made of recycled materials transforming from a pile of plastic bottles into a high-fashion garment. Simple, right?
We spent hours with Pika, trying every prompt we could think of. “A dress made of plastic bottles turning into a couture gown, hyperrealistic, 8k, cinematic.” The results were garbage. We got monstrous plastic-dresses, half-melted mannequins, and gowns that looked like they were from a low-budget sci-fi movie. The team was getting demoralized. The founder even said, “Maybe this tech just isn’t there yet.”
I refused to believe that. I’ve seen what’s possible. The problem wasn’t the tool; it was us. We were thinking like humans, not like machines. That night, I couldn’t sleep. I started digging through the public Pika channels, scraping data, and looking for patterns. That’s when I decided to turn this into a full-blown project. I needed to understand what really works.
My 50,000-Generation Analysis
I focused on a few key things. I looked at the prompt length, the specific keywords used, the use of negative prompts, and camera commands. I also built a simple scoring system to rate the aesthetic quality and coherence of the output. It wasn’t perfect, but with 50,000 data points, the trends became incredibly clear.
What I found surprised me. The most successful prompts weren’t always the longest or the most detailed. They were the most suggestive. They painted a mood and let the AI fill in the blanks. This insight was the key to everything.
Pattern 1: The “Cinematic” Trap
Almost everyone uses the word “cinematic.” It’s the most common keyword I found, appearing in over 60% of the prompts I analyzed. And here’s the kicker: it often makes things worse.
The word has become a catch-all that signals “make this look good.” The AI interprets this with the most generic signifiers of quality: high contrast, saturated colors, and, worst of all, excessive lens flare. It’s the visual equivalent of a student trying to sound smart by using big words they don’t understand.
What works better? Being specific about the kind of cinematic look you want. Instead of just “cinematic,” try referencing a director or a specific film style.
- Instead of:
a man walking in the rain, cinematic - Try:
a man walking in the rain, style of Blade Runner 2049, anamorphic lenses, neon reflections on wet pavement
See the difference? The second prompt gives the AI concrete visual anchors. It’s not just a generic mood; it’s a whole visual language. My analysis showed that prompts referencing specific directors or cinematographers (like “style of Wes Anderson” or “cinematography by Roger Deakins”) had a 45% higher chance of producing a top-tier result than those that just used “cinematic.”
This reminds me of advising startups. The founders who say “we want to be the Uber for X” are thinking too small. The ones who can paint a specific, unique vision of the future are the ones who get my attention—and my investment.
Pattern 2: Negative Prompts Are Your Steering Wheel
Most people treat negative prompts like a janitor. They use them to clean up messes. --no text, --no watermark, --no ugly. This is a huge missed opportunity.
A negative prompt is one of the most powerful creative levers you have. It’s not just for removing things you don’t want; it’s for pushing the AI towards things you do want by eliminating alternative paths.
Think of it this way: if you tell the AI to create a “serene forest,” it has a million ways to interpret that. But if you say a serene forest --no dark, --no scary, --no clutter, you are actively guiding it towards a clean, bright, and peaceful aesthetic. You’re carving away the marble to reveal the statue inside.
I found that the most sophisticated users were using negative prompts to control composition and style:
--no symmetry: Creates more dynamic, off-center compositions.--no photorealism: Pushes the model into a more illustrative or abstract style.--no bright colors: Forces a muted, more atmospheric palette.
In my dataset, prompts that used stylistic negative commands had a significantly higher aesthetic score. It’s the difference between letting the car drive itself and actually steering it where you want to go.
Pattern 3: The Magic of Verbs and Actions
This was perhaps my most surprising finding. The choice of verb in a prompt has an outsized impact on the final video. Pika is a video model, and it thinks in terms of motion. Giving it a strong, clear action is critical.
Weak prompts use passive language: a woman in a field.
Strong prompts use active verbs: a woman *running* through a field, her hair *flowing* in the wind.
But it goes deeper. The more evocative the verb, the better the result. Don’t just say “walking.” Say “striding,” “shuffling,” “marching,” or “wandering.” Each of those words carries a universe of stylistic information.
I saw a prompt that said an old man, remembering his past. The result was a static, boring shot of a man staring into space. The user revised it to an old man, memories *flickering* across his face. The output was breathtaking. It showed a subtle, ghostly superimposition of past events over the man’s expression. The verb “flickering” was the key. It gave the AI a visual metaphor to work with.
This is the art of prompt writing. It’s poetry, not just description. You’re not just listing ingredients; you’re writing a recipe for a feeling.
My Go-To Prompt Structure for Pika
After all this analysis, I’ve developed a new standard structure for my own prompts. It’s not a rigid formula, but it’s a starting point that consistently delivers better results.
[Subject] + [Action Verb] + [Setting] + [Stylistic Reference] + [Composition/Lighting] + [Negative Prompt]
Let’s break it down with an example. Let’s go back to that sustainable fashion dress.
Old Prompt:
A dress made of plastic bottles turning into a couture gown, hyperrealistic, 8k, cinematic.New Prompt:
A translucent dress made of shimmering plastic shards *morphing* into an elegant silk gown, on a runway, style of Alexander McQueen, dramatic uplighting, volumetric shadows, --no mannequin, --no busy background.
Every part of that new prompt is doing a specific job:
- Subject: “A translucent dress made of shimmering plastic shards” (More specific and evocative than just “plastic bottles”)
- Action Verb: “morphing” (A much better action word than “turning into”)
- Setting: “on a runway” (Provides context)
- Stylistic Reference: “style of Alexander McQueen” (A powerful, specific aesthetic)
- Composition/Lighting: “dramatic uplighting, volumetric shadows” (Controls the mood)
- Negative Prompt: “--no mannequin, --no busy background” (Cleans up common failure modes)
When we used this new prompt structure, we got the shot on the third try. It was the centerpiece of the pitch deck, and the company closed their seed round two weeks later. The investors said it was the video that sold them on the vision.
It’s Not About One-Shot Magic
Even with the perfect prompt, you’re not always going to get it right on the first try. The real skill of an AI creator is iteration. It’s about taking the output you get, analyzing what worked and what didn’t, and refining your prompt.
This is the same feedback loop we use when building products. You build an MVP, you get user feedback, and you iterate. You don’t expect to build the final product on day one. Why would AI generation be any different?
I see so many people get frustrated and give up. They try a prompt, it doesn’t work, and they blame the tool. That’s like a programmer blaming the compiler because their code has a bug. The tool is just a mirror reflecting your instructions. If you don’t like the reflection, change the instructions.
My analysis of 50,000 generations taught me that the best AI artists are relentless iterators. They are detectives, constantly searching for clues in the output, refining their approach with every generation. They are playing a game with the machine, learning its quirks and biases, and using them to their advantage.
This is the real secret. It’s not about finding a magic list of keywords. It’s about developing a deep, intuitive understanding of how the model thinks. It’s about being a partner with the AI, not just its boss. The sooner you embrace that, the sooner you’ll start creating things that truly feel like magic.
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 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.
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.