Title: Mastering AI Prompt Engineering: Practical Tips for Midjourney, DALL‑E 3, Stable Diffusion, Sora, RunwayML, ChatGPT, and Claude
Meta: Dive into advanced prompt‑engineering secrets for popular AI models. Learn how to craft precise, high‑quality prompts, use negative instructions, style modifiers, and technical parameters to unlock creative potential.
---
Introduction
Prompt engineering is the craft of guiding AI models to produce the exact output you want. Whether you’re sketching surreal landscapes for Midjourney, generating photorealistic images with DALL‑E 3, looping video transitions in RunwayML, or striking conversations through ChatGPT and Claude, a smart prompt can be the difference between a mediocre result and a masterpiece. This article bundles the most effective techniques, complete with real‑world examples of good versus bad prompts, negative instructions, style keywords, aspect ratios, and quality parameters.
---
1. Understand the Base Language of Each Model
| Model | Core Prompt Syntax | Key Parameters |
|-------|--------------------|----------------|
| Midjourney | Short creative sentences + style tags | –v version, –ar aspect ratio, –q quality |
| DALL‑E 3 | Natural‑language description | size, style, image prompt text |
| Stable Diffusion | Text + negative prompt | --aspect, --cfg_scale, --steps |
| Sora | Text + cinematographic tags | --aspect, --style, --seed |
| RunwayML | Prompt + tool‑specific tags | --style, --resolution, --augments |
| ChatGPT/Claude | Conversational prompt + system instruction | system, temperature, max_tokens |
Knowing these foundations lets you hit the sweet spot from the first shot.
---
2. Start With a Clear Goal
Good Prompt
> “A cyberpunk cityscape at night, neon‑lit skyscrapers reflected in rain‑slick streets, 8K resolution, cinematic lighting.”
Bad Prompt
> “A city with a lot of lights, it’s from the future, super bright.”
Why the first works: It describes atmosphere, lighting, perspective, and resolution—all factors that shape the output. The second is vague and leaves many decisions up to the model.
---
3. Leverage Style Modifiers
Style tags guide the artistic direction. Use them sparingly and deliberately.
| Modifier | Effect | Example |
|----------|--------|---------|
| ::3 | Priority level (Bing image creator) | ultrawide,__shimmering sun::3 |
| in the style of <artist> | Mimic specific artists | in the style of Van Gogh, swirling starscape |
| photorealistic | Emulates real photography | office interior, photorealistic, HDR |
| 8k ultra detail | Emphasizes high resolution | A macro shot of dew on a leaf, 8k ultra detail |
Try to keep modifiers concise; too many can dilute focus and cause the model to wander.
---
4. Use Negative Prompts to Exclude Unwanted Elements
Negative prompts act like “do not” instructions. They’re especially useful for Stable Diffusion and Sora.
Example (Stable Diffusion)
> Prompt: “A serene lake with pine trees.”
> Negative: “no people, no cars, no text”
Example (Sora)
> Prompt: “An aerial shot of a mountain range.”
> Negative: “exclude water bodies, avoid shadows”
Negative prompts work in DALL‑E 3 too, by adding “no” clauses in the description: “A portrait of a person, no glasses”.
---
5. Master Aspect Ratios and Resolution
Aspect ratios determine horizontal to vertical dimensions; resolution dictates pixel density.
Midjourney
- <code>–ar 16:9</code> – wide cinematic
- <code>–ar 9:16</code> – vertical for mobile
- <code>–q 2</code> – twice the rendering time & quality
DALL‑E 3
Use size: 1024x1024 for default square, size: 1200x800 for panoramic. Scaling above 1024x1024 may need size: 2048x2048 in the API.
Stable Diffusion
--aspect 21:9 or --aspect 4:5 are added in the command line. The --steps parameter (usually 50–100) can significantly affect sharpness.
Remember: larger aspect ratios usually require higher steps for clarity.
---
6. Pinpoint the Temporary vs Permanent Attributes
Temporary attributes: mood, color palette, style.
Permanent attributes: composition, key objects, subject.
When modifying prompts, start with the permanent backbone and layer temporary adjectives on top.
Example (Midjourney)
> “A close‑up of a red apple on a marble counter – the scene is saturated with warm amber light – photorealistic, detailed, 4K”
---
7. Practical Tips for Each Model
| Model | Quick Tips |
|-------|------------|
| Midjourney | • Use –v 5 for better detail.
• Combine comma and :: priorities.
• Keep the prompt under 50 tokens. |
| DALL‑E 3 | • Add style: “oil painting” in the JSON.
• Use prompt_strength to blend image prompts.
• Turn off “narrative text” for clean results. |
| Stable Diffusion | • Add --cfg_scale 8 for stricter adherence.
• Apply --seed 42 for reproducibility.
• Cascaded diffusion for higher detail. |
| Sora | • Use cinematographic tags: cinematic lighting, slow motion.
• Hit --style cinematic for low‑res reference. |
| RunwayML | • Try -t “video style transfer” for dynamic textures.
• Add --dropout 0.1 to avoid ghosting. |
| ChatGPT/Claude | • Pre‑preamble: `system: “You are an expert visual