This GPT Image 2 prompt is going insanely viral right now. The full text tells the model to take any attached image and redraw it in the most clumsy scribbly and utterly pathetic way possible. It calls for a white background and the exact look of something drawn in MS Paint with a mouse. The result should stay vaguely similar to the source but also miss in confusing awkward ways. It emphasizes the low quality pixel by pixel feel. Then it ends with a casual whatever just draw it however you want.
The prompt succeeds because it forces a capable image model to produce its worst possible output on purpose. I tested it across a range of source photos and the results stay consistently bad in exactly the requested style. A photo of a busy downtown street came back with buildings reduced to uneven vertical lines that tilt at odd angles. Windows appear as random blue squares floating outside the frames. The road is one thick crooked smear of gray with pixel edges that do not line up. It looks like someone who understood the basic layout tried to draw it freehand with a bad mouse and no undo button. The model committed to the flaws without any smoothing or correction.
A portrait photo produced a face where one eye sits noticeably higher than the other. The hair falls into six or seven thick uneven strokes that ignore any sense of flow. The mouth sits in a half smile that reads as confused rather than intentional. These outputs do not look like stylized art. They look like actual bad drawings that somehow retain just enough of the original subject to remain recognizable. That balance explains much of the sharing. People expect AI images to look polished or at least competent. Forcing the opposite creates outputs that feel surprising and worth passing along.
What stands out is how precisely the model follows directions that earlier systems tended to fight. Older image generators would often improve proportions or clean up lines even when asked to produce low quality work. This version accepts the assignment to be pathetic. The specific references to MS Paint and mouse drawing give it clear anchors from its training data on old software and amateur digital art. The final casual line removes any remaining pressure to make the result coherent or pleasing. The model treats that as license to lean all the way into randomness and error.
I see direct uses for this in certain kinds of work. When building a meme or social asset that must avoid any hint of professional finish this prompt delivers the right tone fast. Marketing teams working on ironic or retro campaigns can generate mockups that look thrown together on an old computer without spending time in actual drawing software. The outputs carry an authenticity that comes from matching the exact aesthetic of hurried amateur work. Internal presentations that need a humorous slide or a visual that signals this is a rough draft rather than final version can benefit too. The time saved compared to opening MS Paint myself and attempting to draw poorly on purpose is real.
This fits the pattern I notice across AI tasks. I do not use the same model or approach for every job. For code I switch tools based on whether the work needs deep codebase understanding or quick generation. The same logic applies here. Most of the time I want image outputs that look sharp and well composed. This prompt adds a reliable way to get the other end of the spectrum when the project demands it. The consistency across different source images means I can generate several variations and select the one that hits the right level of awkward for the goal.
The virality also highlights something useful about prompt structure. Naming the exact tool the specific mistakes and the target level of incompetence produces tighter results than a vague request to make it look bad. The prompt does not just say lower the quality. It describes cursor paths that shake circles that fail to close proportions that mismatch and color that bleeds. Those details map directly to what the model knows about bad mouse drawings from the early days of personal computing. Prompt writers can borrow this approach when they need control over style execution rather than just content.
I ran the prompt on a product shot of a car. The output showed wheels of noticeably different sizes. The body looked like stacked rectangles with sides that refused to stay parallel. Headlights became two yellow dots placed near the right spots while the grille dissolved into an unrecognizable cluster of black lines. The overall shape stayed close enough that you could identify it as that same car but every traditional measure of good illustration was deliberately broken. A landscape image turned hills into wavy lumps with trees reduced to brown sticks topped by green blobs that overlapped in wrong layers. Each test reinforced the same takeaway. The model can produce both ends of the quality scale when instructed clearly.
After months of seeing AI visuals that trend toward perfection many people seem to enjoy outputs that own their defects. This prompt turns limitation into a feature. It gives permission to celebrate the imperfect. That shift influences how I write other prompts now. Instead of battling the model to reach an ideal I sometimes describe the exact type of imperfection the task needs. The white background choice removes any competing elements and keeps every flaw obvious. Blocky edges color banding and off register details stand out sharply against the empty field. Those accumulated choices are what make the images travel so well.
The prompt also works as an informal test of instruction following. Replicating the specific flaws of a twenty year old program requires the model to understand both the source image and the target medium at a detailed level. It does that translation without inserting modern touches like clean gradients or vector sharpness. The generator stays in character for the entire output. I expect more experiments will follow that push image models toward controlled imperfection on purpose. The particular text may stop spreading at some point but the method of asking for bad drawing will remain useful.
In my own projects I reach for image generation when I need fast visuals to illustrate text or explore an idea quickly. This prompt expands the available range. It will not replace design work that needs to look tight and professional. It does earn a regular place for humorous content rough concepts or any situation where the audience should feel the image was assembled rather than designed. The larger point for me is flexibility. Systems that can hit both high quality and low quality targets when asked give creators more range than before. We gain the ability to specify the exact degree of badness a project requires and get output that matches without resistance.
I keep a copy of the full prompt saved for the next time a task calls for something that looks like the work of a distracted person using an old trackball mouse. The results hold up because they avoid the common failure mode of trying too hard to be bad. They simply are bad in a way that still connects back to the source material. That connection is what separates these images from random scribbles. For anyone who works with visuals regularly it represents one more targeted option in the toolbox. The contrast with the model’s normal high capability outputs supplies the surprise that drives the shares. Once you see a few examples it becomes obvious why the prompt caught on so quickly.
The technique aligns with how I approach other parts of AI use. Different jobs need different settings. Budget models suffice for straightforward classification while frontier models handle complex reasoning. Here a single well crafted instruction flips the output style completely. That kind of control matters more than any single benchmark. It lets me match the visual tone to the message without extra steps. I have generated batches of these redraws for test projects and the hit rate is high enough that picking the best one takes little time. The prompt delivers on its promise every time I have used it.