Prompt Writing Basics

1. What is an AI Image Prompt?

An AI image prompt is a text instruction that tells an AI image model what kind of image to generate. Unlike chat prompts to a language model, image prompts are traditionally composed of comma-separated keywords (tags) that each describe a single visual element. The model reads these tags, matches them against concepts it learned from its training data, and synthesizes an image that tries to satisfy all of them at once.

Example: masterpiece, best quality, 1girl, long hair, blue eyes, school uniform

Each tag instructs the AI to include a specific visual element. Well-combined tags produce images closer to your vision; poorly combined ones produce something technically correct but aesthetically wrong. This guide teaches the difference.

2. Basic Prompt Structure

Effective prompts follow this ordering:

step 1Subject and clothing only

step 2+ appearance and expression

step 3+ pose and setting

step 4+ light and framing

Shared prompt

Differs per cut — each cut on the comparison page has it in full.

Negative prompt

nsfw, nude, topless, bottomless, bikini, swimsuit, lingerie, underwear, panties, bra, cleavage, ass focus, breast focus, thigh focus, wet skin, suggestive, school uniform, pleated skirt, blazer, serafuku, child, loli, teenage, bad quality, worst quality, worst detail, sketch, censor, jpeg artifacts, watermark, signature, text, extra limbs, bad hands, bad anatomy, missing fingers, extra digits, deformed, mutated, blurry, lowres

Model

rinFlanimeIllustrious_v30 · euler_ancestral/normal · steps 28 · CFG 5 · 896×1152 · seed 777005

This comparison was measured on a single Stable Diffusion family checkpoint. The same tags can behave differently on NovelAI, Midjourney or another checkpoint — read it as a reference point, not a verdict.

3. Tag Order Matters

Tags placed earlier carry more influence on the output.

Promgrammer's category order is already optimized. Selecting tags top-to-bottom produces well-ordered prompts. You can also drag and drop to reorder.

4. Good vs Bad Prompts

5. Practical Tips

6. Token Budget — Why 15–25 Tags Works Best

AI image models don't count tags — they count tokens. A token is a chunk of text that the model's tokenizer treats as a single unit. Most common tags like masterpiece or 1girl are a single token. Compound or unusual tags such as baroque_costume can split into three or four tokens.

Stable Diffusion 1.5 has a hard 75-token limit per prompt segment. SDXL and newer models allow longer prompts, but practical quality plateaus around 150 tokens. That maps roughly to this guideline:

If you find yourself past 25 tags, try removing synonyms or consolidating two specific tags (long_hair + black_hair) into one compound phrase the model already knows.

7. Before / After — Three Prompt Iterations

The fastest way to understand prompt refinement is to watch the same idea evolve across revisions. Each example below goes Before → After v1 → After v2, with short notes on what changed.

8. Five Common Beginner Mistakes

1. Writing prompts as sentences. "A beautiful girl with long hair standing in a forest" works for Midjourney and DALL·E, but on Stable Diffusion or NovelAI you want comma-separated tags instead. If you're not sure which model expects which style, default to tags — they degrade gracefully.
2. Forgetting commas. long hair blue eyes might be parsed as a single concept. Always separate tags with a comma and a space, even inside category groups.
3. Leaving the negative prompt empty. A good negative prompt is the single biggest quality lever most beginners skip. Start with low_quality, blurry, bad_anatomy, extra_fingers, watermark, text and you'll already see fewer distorted hands and overlayed logos.
4. Dropping quality tags entirely. Even a single masterpiece at position 1 measurably improves fine-tuned models. The cost is one token; the benefit is real.
5. Over-using weight syntax. Putting (tag:1.3) on three or four tags at once cancels itself out — the model just biases the whole prompt and loses specificity. Boost at most one or two critical tags.

9. Frequently Asked Questions