The pattern that changes everything: be specific first
Most people treat the AI prompt like a question to be answered. “Write me a scene.” That’s why the output is generic. You’ve asked a generic question. The AI defaults to average. The fix is simple: don’t ask the AI to write. Ask the AI to write like something specific. Not “write a scene,” but “write a scene in the style of a Safdie brothers thriller, where the tension builds through blocking, not dialogue.” You’ve just eliminated half the generic output space. You’ve given the AI a target that isn’t “average scene,” it’s “Safdie-style scene.” The AI will still miss the mark sometimes, but you’ve made it impossible for it to generate the 75 percent compromise that looks fine but feels hollow.
This works across every tool. “Describe these assets” produces vague descriptions. “Describe these assets the way a cinematographer would—focusing on light quality, shadow direction, and practical light sources” produces usable descriptions. “Generate three script ideas” produces three forgettable premises. “Generate three script ideas in the mode of a genre subversion—each one breaks an expectation of [genre], but stays emotionally true to it” produces something you might actually use. The principle is identical: make the AI’s job smaller and more specific. Don’t ask for quality. Specify what quality means in your context.
How to use examples to train the AI without training it
The second pattern feels like cheating, but it’s legal. Show the AI an example of what you want, then ask it to generate more like it. You don’t need the AI to understand aesthetic theory. You just need it to pattern-match on something concrete you’ve already made or found. If you’re building a library of location descriptions, find one location description you like, paste it into the prompt, and say “Generate nine more location descriptions in this style.” The AI will analyze the example and reproduce its qualities. It won’t be perfect. But it’ll be in the same universe as your reference. This works better than trying to describe the style in words, because style is felt, not explained. An example teaches faster than a manifesto.
For script feedback, this is gold. Instead of asking “Give me character notes,” paste a scene you like and ask: “This scene works because of [specific reason]. Apply that same principle to my scene.” You’re not asking for generic feedback. You’re asking for feedback shaped by a specific example. The AI learns from the pattern in the reference, then applies it to your work. It’s collaborative in a way that generic prompts never are. Director David Fincher talks about showing references instead of describing them. The same principle works with AI tools. A five-second video clip teaches more than a five-paragraph description. A screenshot teaches more than an explanation. Use reference material in your prompts, not as decoration, but as the actual instruction.
Three prompts that work for any creative project
The first meta-prompt is the reverse question. Instead of telling the AI what to do, ask it to tell you what you’re doing. “I’ve written this scene. What emotional beats are actually working here, and what’s their order?” This makes the AI a mirror, not an author. It teaches you to see your own work by having the AI articulate what it sees. Once you understand what you’ve done well, you can repeat it intentionally. This is faster than workshopping with another human because you get feedback instantly, and it’s consistent every time you ask it the same scene. The AI becomes a diagnostic tool, not a creative one. Use it to understand your intuition, not replace it.
The second meta-prompt is the constraints approach. Instead of “make it better,” say “make it work with these constraints.” Give the AI a problem to solve, not a quality bar to hit. “This scene is three pages, but it needs to be one page. Cut it by two-thirds, keeping the emotional core.” Now the AI is doing editorial work, not creative work. It’s solving a structural problem. Constraints produce better outputs than open-ended requests because they give the AI something concrete to aim for. It’s the same reason improvisation rules like “yes, and” work in creative collaboration. The constraint is what makes the work possible.
The third meta-prompt is the diagnosis prompt. “This section doesn’t work. What’s failing, and why?” Let the AI figure out what’s wrong before trying to fix it. Most people skip this step. They ask for a fix without understanding the problem. The AI will bandage the symptom. But if you ask it to diagnose first, then propose fixes, the fixes are often actually useful. Ask: “What assumption is this scene making that the audience might not accept?” or “Where is this dialogue doing work that should be done visually?” The AI becomes a spotter of problems, not a solver of problems. That distinction matters. Problems you understand can be fixed better than problems you’re told to fix.
The one mistake that wastes more time than anything else
You ask the AI for ten options when you really only need two good ones. You’re trying to see the full solution space before committing. But AI quality degrades with quantity. The first option is usually the best option. The second is worse. By the tenth, you’re getting noise. You now have ten mediocre options and need to spend time evaluating all of them. The time you thought you’d save by asking for ten options gets spent narrowing ten options down to two. Ask for two. If both fail, ask for two more. This is slower per batch but faster overall because you’re not wading through garbage. Quality iteration beats broad iteration when you’re working with AI.
The related mistake is over-iterating on a bad output. If the AI’s first attempt is fundamentally wrong, iterating on it won’t fix it. You need to change the prompt. Change the constraint. Change the reference. Change the example. But most people iterate: “That wasn’t quite right, make it a little better.” The AI adjusts by 5 percent. You ask again. It adjusts by another 5 percent. Twenty iterations later, you’ve got output that’s still fundamentally wrong, just slightly more polished. If the first output is off-target, restart. New prompt. New approach. Iteration improves existing directions. It doesn’t fix wrong directions. Save your iteration budget for when the AI is close. Spend your direction-change budget early.
Share your thoughts