I’m going to be honest. Six months ago, if you’d asked me about prompt engineering, I would have shrugged. I’d built and sold two tech companies, invested in over 200 startups, including some of the biggest names in AI like Anthropic and OpenAI, but the nitty-gritty of “prompting” felt like a niche skill for researchers. I was wrong. Dead wrong.
It all started when I was playing around with DALL-E, trying to generate some cover art for my book, Becoming Top 1%. I typed in something simple like, “a book cover about becoming successful.” The results were… laughably bad. Cheesy stock photo-style images of people in suits high-fiving on a mountain top. It was a wake-up call. I realized that these incredibly powerful AI models were not mind readers. The quality of the output was a direct reflection of the quality of the input. My journey (or obsession, as my wife might call it) with prompt engineering began that day.
I didn’t just dip my toes in. I went deep. I’m talking about spending hundreds of hours and thousands of dollars on API calls, running experiments, and meticulously documenting what worked and what didn’t. I approached it like I would any other engineering problem: systemically, with a focus on first principles. And after six months of relentless testing, I’ve distilled my findings into the ten most critical, non-obvious lessons that will hopefully save you from making the same costly mistakes I did.
1. Your AI is a Super-Intern, Not a Mind-Reader
This is the absolute first thing you have to internalize. We see these slick demos from AI companies and start to think these models are sentient. They’re not. Think of the most brilliant, eager, but utterly inexperienced intern you’ve ever met. They can do incredible things, but they need excruciatingly specific instructions. They have no context about your business, your personal taste, or the unspoken nuances of your request.
I remember trying to generate a logo for a portfolio company. I just wrote “a logo for a fintech startup.” The results were a sea of generic dollar signs and circuit board patterns. It was only when I got specific, with things like “a minimalist logo for a mobile-first investment app targeting millennials, using a deep forest green and gold color palette, with a stylized ‘A’ that subtly incorporates a rising arrow”, that I started getting something usable. You have to provide the context the AI lacks. Who is this for? What is the emotional tone? What are the constraints? Be the art director.
2. The “Negative Prompt” is Your Most Powerful Tool
This one feels counter-intuitive, but it’s probably the biggest unlock I found. We’re so focused on telling the AI what we want to see that we forget to tell it what we don’t want to see. Most image generation tools have a dedicated field for negative prompts, and it’s your best friend for refining outputs.
Early on, I was generating images of futuristic cities and kept getting these weird, distorted-looking cars. They looked like they were from a 1990s video game. I tried adding “hyper-realistic cars” to my positive prompt, but it didn't help much. The breakthrough came when I added “low-resolution, video game, distorted, ugly, malformed cars” to the negative prompt. The difference was night and day. The AI suddenly understood the quality I was after by learning what to avoid. The same goes for avoiding clichés. Generating a hero image for a blog post? Add “cliché, stock photo, generic, 3D render” to your negative prompt. It’s like putting up guardrails for the AI’s creativity.
3. Iterate, Iterate, Iterate (and Document Everything)
Nobody gets the perfect prompt on the first try. Nobody. The process is a loop: prompt, generate, analyze, refine, repeat. The key is to be systematic about it. Don’t just randomly change words. Change one variable at a time and see what happens. Is it the verb? The adjective? The order of the words?
I have a massive spreadsheet with hundreds of prompts I’ve tested. Each row has the prompt, the negative prompt, the seed number (if the tool provides it), a link to the output, and a “notes” column where I jot down my observations. It sounds tedious, and it is. But it’s the only way to build a real intuition for how the model “thinks.” For instance, I learned that for DALL-E 3, front-loading the most important concepts in the prompt has a huge impact. For Midjourney, the stylistic parameters are everything. You only learn this through methodical iteration. It’s not art, it’s science. If you want to learn more about how I approach building systems, you can check out my post on building a second brain.
4. Specificity is King, but So is Analogy
I’ve already talked about being specific, but there’s a nuance here. You can be so specific that you corner the AI and stifle its creativity. Sometimes, the best way to communicate a complex idea is through analogy. Instead of describing every single detail of a character’s face, you might say, “a wise old wizard who looks like a retired rockstar.” The AI has a rich, deep understanding of who a “retired rockstar” is—the leathered skin, the mischievous glint in the eye, the slightly unkempt hair. You’re tapping into the model’s vast network of associations.
I was trying to create an image for a post about startup pivots. I described a complex scene of a half-built bridge changing direction mid-air. The results were a mess. Then I tried a different approach: “A massive, ancient tree that has been struck by lightning, but from the blackened trunk, a bright new branch is growing in a completely different direction, covered in bright green leaves.” Boom. The analogy communicated the concept of destruction and rebirth far more effectively than my literal description ever could.
5. Master the “Weights” and “Seeds”
This is where you start to go from amateur to pro. Many advanced AI tools allow you to add “weights” to different words in your prompt. This tells the model how much importance to place on each concept. For example, in the prompt “a red car in a city,” you could weight “red” higher to ensure the car is the dominant visual element. It’s often written like this: (red:1.5) car in a city. That 1.5 tells the AI to pay 50% more attention to the word “red.” Learning to manipulate these weights is how you gain fine-grained control over your compositions.
Similarly, the “seed” is a number that controls the initial noise pattern the AI starts with. If you use the same prompt with the same seed, you’ll get the exact same image. This is incredibly powerful for iteration. When you get an image that’s almost perfect, you don’t have to start from scratch. You can lock in the seed and then just tweak the prompt, knowing that the underlying composition will remain stable. It’s the key to consistent and predictable results.
6. The Power of “Shot” and “Lighting” Keywords
If you want your images to look less like AI generations and more like professional photographs or film stills, you need to start thinking like a cinematographer. The two most impactful sets of keywords you can learn are related to camera shots and lighting. Don’t just say “a picture of a woman.” Say “a full-body shot of a woman” or “a close-up, intimate portrait of a woman.” Use terms like “wide shot,” “Dutch angle,” “point-of-view shot.”
Lighting is even more important. It’s the difference between a flat, boring image and a dramatic, emotional one. Learn the language of light. Is it “soft, diffused morning light”? “Harsh, direct midday sun”? “Dramatic Rembrandt lighting”? “Cinematic, volumetric lighting”? “Neon, cyberpunk lighting”? Each of these phrases unlocks a completely different mood and aesthetic. I once spent a whole day just generating the same simple scene, a coffee cup on a table, but with dozens of different lighting keywords. It was a more valuable education than any tutorial I’ve watched.
7. Don’t Underestimate the “Weird” Prompts
Sometimes, the most interesting results come from the most unexpected inputs. Throwing a strange or abstract concept into your prompt can break the AI out of its default patterns and lead to something truly original. I’m talking about things like “a portrait of a CEO, but made of liquid data” or “a cityscape that is also a musical instrument.”
This is a great way to discover new styles. One of my favorite techniques is to combine a very concrete subject with a completely unrelated artistic style. For example, “a photorealistic portrait of a medieval knight, in the style of a 1980s airbrushed van art.” The clash of concepts forces the AI to get creative and synthesize something new. It’s a bit like being a mad scientist, but the results can be genius. It reminds me of the early days of startups, where the craziest ideas sometimes turn out to be the best. You can read more about my thoughts on this in my post on why you should build a startup in a down market.
8. The Model Matters. A Lot.
Not all AI models are created equal. DALL-E 3 is fantastic at understanding natural language and creating illustrative, almost whimsical images. Midjourney, on the other hand, is the king of aesthetic cohesion and artistic, moody compositions. Stable Diffusion, especially when run locally, offers the ultimate control and customization, but has a much steeper learning curve. You need to understand the strengths and weaknesses of each.
I have different tools for different jobs. If I need a quick, clean illustration for a blog post, I’ll often start with DALL-E. If I’m trying to create a truly stunning, portfolio-worthy piece of art, I’ll spend the time crafting the perfect prompt in Midjourney. And if I need to do something highly specific, like generating a consistent character across multiple scenes, I’ll turn to Stable Diffusion and its ecosystem of plugins. Don’t be a monogamist with your AI tools. Play the field.
9. Context is Your Secret Weapon for Text Generation
Prompt engineering isn’t just for images. It’s just as, if not more, important for large language models (LLMs) like GPT-4. The same principles apply. The model is an intern. It needs context. But with text, the “context” is even more critical. Before you ask the AI to write something, you need to load its “brain” with the right information.
This is called providing a “meta-prompt” or “pre-prompt.” You start by giving the AI a role, a persona. For example, you could say: “You are a seasoned Silicon Valley investor with two successful exits. You are writing a blog post for other entrepreneurs. Your tone should be direct, opinionated, and a little bit informal. You use short, punchy sentences and are not afraid to be controversial.” By giving it this persona before you give it the actual writing task, you’ve framed the entire interaction. The results are dramatically better than just saying “write a blog post about X.”
10. The Future is “Prompt Chaining”
This is the final lesson, and it’s the one that has me the most excited. The most advanced workflows don’t rely on a single, perfect prompt. They rely on a sequence of prompts, where the output of one step becomes the input for the next. This is “prompt chaining,” and it’s how you can automate complex, multi-step tasks.
For example, instead of trying to write a whole blog post in one go, you could have a chain of prompts:
- Prompt 1: “Generate 10 potential titles for a blog post about the lessons learned from prompt engineering.”
- Prompt 2 (using the best title from step 1): “Create a detailed, seven-part outline for a blog post titled ‘10 Lessons I Learned About Prompt Engineering After Spending 6 Months’.”
- Prompt 3 (for each section of the outline): “Write a 300-word section for the part of the blog post titled ‘Your AI is a Super-Intern, Not a Mind-Reader,’ using a personal anecdote about trying to generate a logo.”
- Prompt 4: “Combine all the sections and write a compelling introduction and conclusion.”
This is a simplified example, but you can see the power. You’re breaking down a complex creative task into a series of smaller, more manageable steps, using the AI as a partner at each stage. This is the future of knowledge work, and it’s all built on the foundation of solid prompt engineering.
It’s Your Turn
Look, I get it. This can feel overwhelming. Six months ago, I was in the same boat. But the truth is, you don’t need to be a machine learning PhD to get great results from these tools. You just need to be curious, systematic, and willing to learn the language that these AIs understand.
I’ve only scratched the surface here. The field is moving at a dizzying pace, and the techniques that work today might be obsolete tomorrow. But the underlying principles of specificity, iteration, and context will remain. My advice is to pick one tool, whether it’s DALL-E, Midjourney, or even just the ChatGPT interface, and commit to spending a few hours a week just playing. Don’t have a goal. Just experiment. Follow your curiosity. That’s how I learned, and I’m willing to bet it’s how you’ll learn too. Now, what are you going to create?
Frequently Asked Questions
How were these items selected?
Each item on this list comes from direct experience, either from building my own companies or from patterns I've observed across the 200+ startups I've invested in. I prioritize practical, actionable items over theoretical concepts.
Which item on this list has the highest impact?
It depends on your stage and context, but in my experience, the items near the top of the list tend to have the broadest applicability. That said, sometimes the less obvious items create the biggest breakthroughs for specific situations.
Are these recommendations still relevant in 2026?
Absolutely. While specific tools and tactics change, the underlying principles remain consistent. I update my thinking regularly based on what I'm seeing in the market and across my portfolio companies.