Danbooru prompts, no "best quality" type tags, no inpainting. Text to image and then upscaler.
("Hufflepuff" often generates hairclips… :-))