ai-image-prompting

Overview

AI image generation is a design medium with its own grammar: the prompt is your art direction, and vague prompts produce vague images. Skilled prompting combines precise subject description, style vocabulary, compositional direction, and technical parameters — then iterates deliberately. This skill teaches structured prompting that produces consistent, art-directable results.

When to use

• Generating concept art, illustrations, or marketing visuals with AI

• Creating consistent character or style outputs across images

• Improving AI images that look generic or off-brief

• Building prompt libraries and reusable style templates

• Directing AI imagery for brand-aligned content

Core concepts

• • • Prompt anatomy. Subject (what) + action/context (doing what, where) + style (how it looks) + composition/lighting (framing, mood) + technical (aspect ratio, quality params). Missing layers produce generic output.
• • • Style vocabulary matters. "Cinematic" is vague; "35mm film still, shallow depth of field, golden-hour rim light, muted teal-and-orange grade" is direction. Build a personal lexicon of style terms that work.
• • • Specificity beats adjectives. "A cozy coffee shop" → "a corner café at dusk, rain-streaked windows, warm Edison bulbs, a barista pouring latte art, worn leather chairs." Concrete nouns outperform abstract adjectives.
• • • Negative prompting and constraints. State what to exclude (text, watermark, extra limbs, photorealistic when you want illustration). Constraints ("single subject, centered, plain background") prevent the model's worst habits.
• • • Seeds and consistency. For series work: lock the seed, reuse style blocks verbatim, and describe recurring elements identically each time. Consistency comes from repetition, not luck.
• • • The 70/30 rule. AI gets you 70% of the way fast; the last 30% (hands, text, brand accuracy, precise composition) usually needs human editing, inpainting, or compositing. Plan for the finish, don't expect it from the prompt.

Practical workflow

1. 1. 1. Define the brief. What's the image for, what must it communicate, what style, what dimensions? Write this before prompting — the brief disciplines the iteration.
2. 2. 2. Draft the structured prompt. Build all five layers (subject, context, style, composition/lighting, technical). Start with your best guess at full specificity.
3. 3. 3. Generate variations. Run 4-8 variations of the first prompt. Don't judge individual images — judge which direction is closest, then iterate on that one.
4. 4. 4. Iterate one variable at a time. Change the style term OR the composition OR the subject detail — not all three. Systematic iteration converges; random tweaking doesn't.
5. 5. 5. Fix with targeted tools. Wrong hands? Inpaint/regenerate the region. Wrong text? Composite real typography over it. Close-but-not-quite? Image-to-image with a tight prompt and low denoising.
6. 6. 6. Build reusable templates. When a style works, save the style block as a template: "[STYLE]
   • [SUBJECT] + [COMPOSITION]." Templates turn one-off wins into a production system.
7. 7. 7. Finish like a designer. Color grade for consistency, composite brand elements and real type, retouch artifacts. The AI output is raw material; the final image is designed.

Common pitfalls

• • • One-shot expectations. Judging the medium by the first generation. Prompting is iterative — professionals run dozens of generations per final image.
• • • Adjective soup. "Beautiful stunning amazing ultra-detailed 8k masterpiece" — empty intensifiers the model mostly ignores. Replace with concrete visual description.
• • • Ignoring aspect ratio. Generating square and cropping to banner. Compose for the final ratio from the start — composition doesn't survive aggressive crops.
• • • Text in images. AI mangles text reliably. Plan to add all typography in post — never prompt critical text into the generation.
• • • Style inconsistency across a series. Slightly different style terms each prompt produce a Frankenstein set. Lock the style block verbatim.
• • • No disclosure where it matters. Editorial, journalistic, and some commercial contexts require AI disclosure. Know the rules for your use case — and never present AI imagery as photography of real events or people.