After testing 200 images, Brother Long condensed the ComfyUI prompt rules into 6 guidelines.

At first, I thought prompts were just a bunch of keywords—"anime, manga, cute, JK uniform, long hair" randomly thrown together, but the output was always a bit off from what I wanted.
Later I realized: the model doesn't understand natural language, only keywords + probability weights.
❶ Weight syntax, one rule is enough to remember.
(Keyword:Number) — 1.2 for emphasis, 0.8 for de-emphasis, 1.0 default. Nesting not exceeding 3 layers (beyond that, overfitting the image).
Minimal syntax: (Keyword++) = 1.2, [Keyword] = 0.9.
❷ Four-layer golden order
Subject → Environment → Art Style → Quality.
The model pays the most attention to the first 20 tokens; core words must be placed at the front.
1 girl, long pink hair, smiling, kimono,
cherry blossom garden, soft sunset light,
anime style, cel shading, pastel palette,
masterpiece, best quality, 4k, intricate details
❸ Don't use Chinese prompts.
SD1.5/SDXL training data are mainly in English; Chinese tags have very low recognition rates.
English prompts + Chinese annotations are the optimal solution.
❹ Universal negative prompt template (enough for 90% of scenes)
low quality, bad anatomy, extra fingers, deformed eyes,
blurry, ugly, duplicate, watermark, signature
Additional for anime: nsfw, monochrome, flat colors
Additional for realism: plastic skin, wax face, overexposed, noise
❺ Two style vocabularies
- Anime: anime screencap, cel shading, vibrant colors, detailed background
- Realistic: photorealistic, cinematic lighting, 85mm, subsurface scattering, film grain
❻ Troubleshooting image issues (for emergency)
- Irrelevant → mention CFG 7-8
- Overexposure of single element → reduce core word weight / post-processing
- Multiple characters stuck together → rewrite four-layer structure
- Repeated objects → change seed + add "repeat" to negative words
Six rules are simple and copy-paste ready. Whether they’re worth it or not is hard to say; after testing 200 images, I finally stabilized the output.