prompt tweak – An AI contributor tested a simple workflow: ask a chatbot to write the full, generator-ready image prompt for you. Using that approach, Gemini and ChatGPT produced markedly better results for generating a sunflower made of sheet metal in a pencil-drawing style
A sunflower made of sheet metal sounds like the kind of idea you can type into an image generator and hope for the best. But in practice, the difference between a decent result and a genuinely usable one can come down to something small: who writes the prompt—the user, or the chatbot.
Today’s “Prompt of the Day” comes from ZDNET contributor Lance Whitney. who recommends a workflow for people struggling to get consistent outcomes from AI image generators. Instead of crafting a long. fragile instruction yourself. you ask a chatbot to design the exact query for the generator you’re using.
Whitney’s example is specific. The starting point is simple: “I would like to create an image of a sunflower made of sheet metal in a pencil drawing style.” Then the instruction changes. You ask the chatbot to generate a full prompt you can use with a specific tool—either Nano Banana or ChatGPT Images.
The method is straightforward. Whitney explains: you supply the basic details of what you want. and then ask the AI to generate a full prompt based on those details. He also points to a practical benefit that matters when generators are picky: when chatbots draft their own prompts. they’re less likely to include wording that the image generator might flag or refuse. The result is a smoother querying process.
To see how well the approach worked, the test was run with Gemini and ChatGPT—paired with their corresponding image generators, Nano Banana and ChatGPT Images.
In the Gemini test. the prompt returned a pencil drawing described in detailed. material-specific terms: a pencil drawing of a metal sunflower where the petals. leaves. and stem are “composed of sculpted pieces of metal sheet.” It also calls out texture and imperfections—visible welds. visible pencil lines. and shading designed to emphasize the three-dimensional nature of the sculpture. The description places the flower in a garden, with other blurry plants in the background.
The test notes that Gemini suggested more detail than the user would have added on their own prompt. Even so, when that more elaborate description was used in Nano Banana, it produced a “decent image” of a metal-ish sunflower.
The second test—ChatGPT paired with ChatGPT Images—followed the same philosophy. but the returned prompt read like it was built for a very specific kind of output. ChatGPT’s generated prompt describes “a detailed pencil drawing of a sunflower constructed from sheet metal. ” emphasizing thin. slightly curved metal plates. visible seams. rivets. and subtle bends. It also spells out a metal-and-organic center using “layered metal textures resembling industrial components. ” plus a metallic stem and leaves with etched vein details and light wear.
In addition to the subject. ChatGPT’s prompt gets prescriptive about the drawing itself: it asks for a “realistic graphite pencil style” with “fine linework. cross-hatching. and soft shading. ” plus “high contrast between light and shadow” to emphasize texture and depth. It also specifies a minimal background—“white or lightly sketched”—so the metallic sunflower stays the focus. The result described from this test was successful: the user says the image was “very successful.”.
Whitney’s advice doesn’t stop at the first draft. If the prompt a chatbot generates is too long, he advises asking for a shorter option. That shorter version, he says, usually works.
The broader takeaway is also blunt: prompt results vary across AI tools. The article notes that it’s not a one-and-done skill—companies update and replace models, and those changes can affect what the same prompt returns.
There are practical guardrails too. If someone is creating AI images for work, they should check their company’s policy on AI-generated content. The reason is simple and unresolved in law: “With copyright law surrounding AI image generators still in flux. ” many companies may not allow any use of AI image generators. For mockups—like internal presentations—the piece suggests that should generally be fine.
And there’s the privacy side. Using chatbots means sharing personal data with AI companies every time you use a chatbot or other AI tool. The guidance included in the piece points readers to ways to reduce exposure by adjusting settings for certain chatbots. It also mentions privacy-first options—potentially encrypted or otherwise privacy-focused chatbots—even if they may not offer the same features or capabilities as more popular tools.
In the end, the most effective change in this experiment wasn’t switching generators or rewriting the idea of a sunflower made of sheet metal. It was handing the prompt-writing job to the chatbot—then refining length only if it got too heavy to use.
AI image generation ChatGPT Gemini Nano Banana ChatGPT Images prompts Lance Whitney privacy copyright AI tools
So basically the AI writes the instructions for the AI. Wild.
Wait, I thought you were supposed to just type what you want and it works. But now you gotta ask Gemini to rewrite the prompt too? That feels like extra steps to me.
Maybe this is why my pictures always look weird. I just figured the generator was broken or whatever. If you let the chatbot draft the prompt then it won’t get flagged?? So it’s like tricking the filter by wording it nicer lol
This sunflower made of sheet metal thing sounds like a meme not an actual use case. Also aren’t these models basically just copying each other’s prompts anyway? Like you ask ChatGPT to make a prompt for the image tool, then it gives you something slightly different… and suddenly it’s “markedly better.” Sure.