Prompting for media: the transferable skill

This is the most transferable, longest-lasting skill in the course: how to prompt a media model well. The tools change; this craft doesn't. And it's different from prompting a chatbot — you're describing a scene, layering visual information the model composes.

A tool-agnostic structure that works everywhere — think of it as stacking layers (coverage matters more than exact wording):

1. Subject — concrete detail: materials, textures, clothing, expression, age. "A weathered fisherman" beats "a man."
2. Action / pose — what they're doing.
3. Environment — where, and what's around them.
4. Composition / framing — shot type and angle: close-up, wide shot, low angle, overhead, shallow depth of field.
5. Lighting — the single biggest quality lever: source, direction, quality, time of day. "Soft golden-hour side light" transforms an image.
6. Camera / lens language — "85mm portrait lens," "shot on film," "shallow depth of field." This steers photorealism far more reliably than tags like "8K, ultra-detailed."
7. Style & medium — photograph, oil painting, 3D render, anime.
8. Mood & color — the emotional and palette direction.

The important 2026 split: classic diffusion tools (Midjourney, Stable Diffusion) reward dense, comma-separated keyword stacks; LLM-native tools (gpt-image-1, Gemini) reward clear natural-language descriptions and explicit constraints stated in plain English ("make sure there are exactly three apples"). Learn to feel which kind of tool you're in.

Key mechanics you'll use constantly:

• Aspect ratio — set it deliberately; it changes composition, not just crop.
• Negative prompts — list what to exclude (classic tools). Newer LLM-native tools have no negative field — you just say "without a hat" in the prompt.
• Iterate deliberately — your first prompt is a draft. Change one thing at a time (hold the seed) so you learn what each word does, rather than rewriting everything and starting over.

The mindset: you're a director giving specific, visual direction — not typing keywords and hoping. Master this layered, iterative approach and every tool in your stack produces dramatically better output, because they all run on the same underlying craft.