ChatGPT
Perplexity
Claude
Gemini
Sieh dir die echten Beispiel-Prompts an Google DeepMind hat früher Nano Banana Pro vorgestellt: „Mach aus dieser Szene Nacht.“ „Fokussiere auf die Blumen.“ „Ändere das Seitenverhältnis auf 1:1, indem du den Hintergrund reduzierst. Der Charakter bleibt exakt in seiner aktuellen Position fixiert.“ Keine davon liest sich wie die dichten, vollgepackten Prompts, die die Leute die letzten zwei Jahre lang zu schreiben gelernt haben. Sie lesen sich wie Notizen, die du einem Mitarbeiter hinterlässt, der schon weiß, woran du arbeitest, und nur die nächste Anweisung braucht.
That's not an accident of phrasing. It's the actual design of the tool. Nano Banana Pro is built to hold an image in an ongoing exchange: you don't redescribe the whole scene to change the lighting, you just say what to change and it edits the frame you already have. OpenAI's newest release works the same way, just from the chat side rather than the canvas side. Instead of writing one exhaustive prompt and hoping, you can say "make the sky more orange" or "add a coffee cup to the left," and the model modifies the existing image instead of generating a new one from scratch, closer to giving direction to a designer than typing a search query. When OpenAI shipped that generation of the model, ChatGPT Images 2.0, on April 21, es setzte sich sofort in jeder Kategorie der Image-Arena-Bestenliste an die Spitze – mit dem größten je dort gemessenen Vorsprung, was einiges darüber aussagt, wo die Branche entschieden hat, dass der eigentliche Wettbewerb jetzt stattfindet: nicht rein bei der Bildqualität des ersten Versuchs, sondern darin, wie gut ein Modell über mehrere Bearbeitungen hinweg im Gespräch bleibt, ohne den Faden zu verlieren.
That shift changes what the actual skill is. Prompt engineering, as a discipline, grew up around the idea that a prompt was a kind of incantation: get the wording, the modifiers, the syntax exactly right, and the single generation you got back was the whole game. A lot of the early advice reflected that, elaborate templates, keyword stacking, the sense that there was one correct spell for any given image. None of that disappears entirely, but it stops being the center of the skill once the tool expects you to keep talking. Die Frage ist nicht mehr nur „Was sage ich zuerst?“ Es geht um „Was ändere ich als Nächstes – und woran erkenne ich, wann ich aufhören sollte?“
That second question turns out to be the harder one, and it's a more familiar kind of hard. It's the same judgment a photographer uses deciding whether a shot needs one more adjustment or is already right, the same instinct an editor uses knowing which note to give first and which ones can wait. Directing a back-and-forth well means noticing what's actually wrong with a draft rather than everything that could theoretically be different about it, and it means knowing your own taste well enough to recognize the version that's actually finished instead of just the version that's different from the last one.
None of this makes the initial prompt irrelevant. A clear starting instruction still saves you several rounds of correction later, the same way a clear brief saves a designer several rounds of revision. But the center of gravity has moved. The people getting the most out of these tools right now aren't the ones with the cleverest opening prompt. They're the ones who know how to keep a conversation going toward something specific, correcting course a little at a time, without losing track of what they were tatsächlich überhaupt erstellen willst.