It's strange when you think about it. You can describe exactly what you want to an AI image generator — the subject, the setting, the mood — and still get back something that looks like a confused stock photo from 2009. I've stared at plenty of those results, wondering what went wrong. And almost every time, the problem wasn't the tool. It was what I told the tool to do.
What I've realized after generating thousands of images across different platforms is that writing prompts isn't really about describing things. It's about thinking in a way that the model understands. That sounds obvious, but it's easy to forget when you're mid-session, frustrated, and just typing words hoping something sticks. Most bad prompts come from habits we don't realize we have — little assumptions about how the AI "should" work, rather than how it actually works.
Here are the mistakes I see over and over again, including ones I still catch myself making.
1. Assuming the AI knows what you're not saying
This is probably the most common one. You write a prompt like "a warrior in a forest" and expect something epic. Instead you get a generic medieval-ish person standing awkwardly between two trees in flat midday lighting. The problem isn't the model. It's that you gave it almost nothing to work with, so it filled in every gap with the statistical average of what "warrior" and "forest" look like across its training data. And that average is boring.
What I've learned is that the AI doesn't infer intention. It doesn't know you wanted moody lighting or a specific cultural reference unless you say so. The more you leave unsaid, the more you hand control over to a mathematical average. Sometimes that works out, but usually it doesn't.
You don't need to write a novel. But you do need to pin down the variables that actually matter to you. If the lighting matters, say it. If the composition style matters, name it. The model isn't being lazy — it just genuinely doesn't know what's in your head.
2. Overloading the prompt with conflicting details
The opposite mistake happens just as often. Someone gets ambitious and writes something like "a photorealistic oil painting in the style of Van Gogh but also like cyberpunk anime with film grain and sharp focus, 8k, HDR, cinematic lighting." I've done this. The result is usually a visual argument — different aesthetic directions pulling the image apart until nothing looks intentional.
What's happening under the hood is that each descriptor pulls the generation in a different direction. "Photorealistic" pushes toward real-world textures and lighting behavior. "Oil painting" pushes toward visible brushstrokes and pigment mixing. "Anime" pushes toward cel shading and simplified features. The model tries to satisfy all of them at once, and you get a compromise that satisfies none.
A better approach is to pick one dominant aesthetic lane and use other descriptors as subtle seasoning rather than competing main ingredients. If the image needs to feel like a painting, lead with that and let everything else support it, not fight it.
3. Using "vibe words" without visual meaning
Words like "epic," "beautiful," "stunning," "breathtaking." We all use them. I still catch myself typing "masterpiece" as if the AI has some internal quality switch that activates when you compliment it. It doesn't. These words are opinions, not visual instructions. They don't tell the model anything about what should actually appear in the image.
What does "epic" look like in terms of composition? Probably a low angle looking up, dramatic scale, strong contrast, maybe backlighting. Why not just say those things instead? "Beautiful" could mean soft lighting and harmonious colors, or it could mean nothing at all depending on context. The AI doesn't have taste. It only has statistical associations, and vague praise words have weak, inconsistent associations.
I'm not saying you can never use these words. Sometimes they tilt things in a useful direction, especially if a model has been fine-tuned on aesthetic ratings. But if you're relying on them to carry the prompt, you're essentially hoping the model guesses your taste correctly. That's a gamble, not a skill.
4. Getting the detail-to-freedom ratio wrong
This one took me a long time to articulate. There's a sweet spot between too little detail and too much, and it shifts depending on what you're trying to create. If you want a specific, controlled outcome — a product shot, a character with exact features — you need high specificity. But if you want something creative or surprising, over-specifying can suffocate the model's ability to generate interesting results.
I think of it like directing a human artist. If you say "draw a tree," you get their default tree. If you describe every branch, you get exactly what you asked for but maybe nothing beyond it. The best results often come from being specific about the things that matter and loose about the things that don't. "A gnarled oak in late autumn, low evening light, detailed bark texture" tells the model what's important while leaving room for it to make compositional decisions you might not have thought of.
The mistake is treating every element with equal importance. Some things deserve precise language. Other things can be left open so the model can do what it's surprisingly good at — finding visual solutions you wouldn't have specified.
5. Ignoring the model's known weaknesses
Every image model has things it struggles with. Hands are the famous example, but there are others that don't get talked about as much. Complex spatial relationships between multiple objects. Text within images. Specific counts of things (try getting exactly seven apples on a table reliably). Fine details on distant faces in crowd scenes.
Ignoring these isn't optimistic — it's just setting yourself up for frustration. If you prompt for a scene that relies heavily on something the model is bad at, you're going to burn credits fighting an uphill battle. Sometimes the smarter move is to design around the limitation. If you need legible text in an image, plan to add it in post. If you need precise hand positions, consider describing the pose in a way that minimizes hand visibility or simplifies the gesture.
This isn't about lowering your standards. It's about knowing where to spend your effort. Fighting the model's fundamental architecture is usually a waste of time. Working with its strengths gets you further, faster.
6. Relying on the same prompt structure every time
It's tempting to develop a formula. Subject, setting, lighting, style, quality tags, done. That works for a certain type of image, but it also trains you into a creative rut. I started noticing this when all my images began feeling like variations on the same thing, even when the subjects were different. The structure of the prompt was imposing a kind of sameness.
Different concepts need different approaches. Sometimes the most important thing is the emotional tone, and you should lead with that. Sometimes a reference to a specific artist or movement does more heavy lifting than a paragraph of description. Sometimes you want to describe the image as if narrating a moment from a film rather than listing attributes.
Switching up how you write keeps your results from getting stale. It also forces you to think about what actually matters for each image, rather than filling in the same template on autopilot.
7. Forgetting about composition entirely
Most prompt advice focuses on subjects and styles. Composition gets ignored, which is strange because it's often what separates an interesting image from a forgettable one. "A castle on a hill" could be shot from a drone a mile away or from the base of the hill looking up. Those are completely different images.
I started getting much better results when I began including basic compositional cues: camera angle, distance from subject, framing, what's in the foreground versus background. Even simple words like "low angle," "overhead view," "close-up," "wide shot" make an enormous difference. You don't need to know film terminology. Just think about where the viewer is standing and what they're paying attention to.
The model can't read your mind about perspective. If you don't specify, it defaults to whatever was most common in similar training images. That's often a medium-distance eye-level shot that's perfectly fine and perfectly unmemorable.
8. Treating negative prompts as a cleanup tool rather than a creative tool
Negative prompts — telling the model what you don't want — are often used like a spam filter. "No extra fingers, no blur, no watermark." That's fine as far as it goes, but it's a limited way to think about them. Negative prompts can actively shape the aesthetic of an image, not just clean up errors.
If you're going for a specific look, telling the model what to avoid can be as powerful as telling it what to include. "No warm colors" pushes the palette in a direction. "No intricate background detail" keeps focus on the subject. "No soft lighting" steers toward drama and contrast. These aren't just fixes — they're creative decisions.
The mistake is treating negatives as an afterthought. They're part of the prompt's overall direction, and they deserve the same intentionality as the positive description.
9. Not iterating with purpose
There's a difference between iterating and just guessing. I used to tweak prompts randomly — change a word here, add a style reference there — without really tracking what each change did. That's slow learning. What works better is changing one thing at a time and observing the result. Same prompt, different lighting word. Same prompt, different composition angle. You start to build an intuition for what each lever actually does.
The people who get consistently good results aren't the ones who write perfect first prompts. They're the ones who know how to read what went wrong and adjust the right variable. That skill only comes from deliberate iteration, not from hoping the next random tweak will work.
10. Blaming the model instead of the instruction
This is the one that probably holds people back the most. When an image doesn't turn out, it's easy to think "this model isn't good at that" or "AI can't do what I want." Sometimes that's true — there are genuine limitations. But a surprising amount of the time, the issue is in the prompt, not the capability of the system.
I've seen people declare that a model "can't do" a certain style or subject, then watch someone else produce exactly that with a differently structured prompt. The difference wasn't the tool or the skill level of the user in some abstract sense. It was that one person had figured out how to communicate what they wanted in language the model responded to.
Assuming the problem is always in your prompt isn't quite right either. Some things genuinely don't work well yet. But defaulting to "the model failed" cuts off the curiosity that leads to getting better. Defaulting to "what could I have said differently" keeps you improving.
Most of these mistakes boil down to the same thing: forgetting that prompting is communication, not command. The AI isn't a reluctant employee you need to manage. It's more like a collaborator that takes everything literally, doesn't ask clarifying questions, and has no shared context with you unless you provide it. That's an unusual kind of relationship, and it takes time to develop an intuition for it.
What changed things for me was shifting from writing prompts to writing descriptions that think like the model thinks. Not in some technical sense — I don't understand the architecture deeply — but in a practical sense. What information does this system actually need to move in the direction I want? What am I saying that's useless? What am I not saying that I'm assuming?
The prompts that work aren't necessarily the longest or the most clever. They're the ones where every word earns its place, where the important visual decisions are specified and the unimportant ones are left alone, and where the writer has actually thought about what they want rather than just typing until it feels like enough.
That's a skill like any other. It's learnable. And the fastest way to learn it is to notice which of these mistakes you're making and fix them one at a time.