Prompting like an art director — one anatomy, four dialects — step 1 of 8

The prompt is a creative brief — write it like one

You already have the hardest skill in image prompting. You've spent your career taking "make it pop" from a client and turning it into a spec a junior can execute: subject, medium, composition, palette, and the three things that absolutely must not appear. That translation is prompting. The only new part is who's on the other end.

A production prompt has five sections, and they map one-to-one onto a brief you'd hand a human:

1. Subject — what's in the frame. Not "coffee stuff." A rooftop espresso bar at dusk.
2. Medium — what kind of artifact this is. Flat vector illustration. Risograph print. Photograph. This single token moves the output more than any other.
3. Style — the execution spec. Stroke weight, texture, print process, camera language. 2px uniform stroke, isometric.
4. Palette — named colors, ideally your brand's. Terracotta and deep teal on off-white.
5. Negatives — the do-not-appear list. No gradients, no photorealism, no watermark.

Skip a section and the model doesn't leave it blank — it fills it with the statistical average of everything it's seen. The average of all posters is the poster you'd be embarrassed to present.

Same anatomy, four dialects

The four tools you'll actually meet in client work all consume this anatomy, but each speaks its own dialect:

• Midjourney — the taste engine. Prose plus parameters: --ar 4:5 for aspect, --no for negatives, --sref for style references, --seed for repeatability. Wins on aesthetic instinct; current Midjourney's quoted-text rendering finally made lettering legible, but it still trails the text specialist for letter-perfect copy.
• Flux — the literalist. Full descriptive prose, no flag syntax; it follows compositional instructions ("subject in the lower-left third") more faithfully than the taste engines. Photorealism is its home turf.
• Ideogram — the typographer. The one family built for text-in-image. Put your exact copy in quotes and it usually sets it.
• nano-banana — the production intern. Google's speed-optimized image family: conversational, strong at instruction-based edits, and cheap enough to generate fifty candidates and keep three. (Which tier is cheapest rotates; the batch-workhorse role doesn't.)

Version numbers on all four rotate every few months; the anatomy doesn't. Treat every tool name, version, flag, and price in this chapter as an illustrative mid-2026 snapshot, not a curriculum-verified fact — glance at the vendor's docs before you bill a client for any of them. The skill this chapter teaches is the anatomy, which outlives whichever model wins next quarter.

(If you've done chapter 27, the AI image generation deep-dive: same territory, different altitude. Its "eight knobs" are the general builder's checklist for any image prompt; this five-section brief is the same anatomy folded into the shape a designer already thinks in, with a brand on the line. Neither contradicts the other — the brief is the knobs, pre-sorted.)

The rest of this lesson drills the anatomy until it's reflex, then the checkpoint has you compile three briefs into their tools' dialects — which is exactly what a busy Tuesday looks like.