A Prompt-Iteration Method

Hitting generate again feels like progress because the image changes. But a fresh roll changes everything at once - a new seed, a new composition, new details - so even when you land something better you have no idea why, and you can't get back to it or build on it. You've bought one image and learned nothing transferable.

Worse, rerolling is seductive. The variety is fun, the occasional gem is real, and an hour vanishes into a pile of pretty near-misses none of which is quite right and none of which you can improve. This is the single biggest time-sink in AI design work, and escaping it is the whole point of this lesson.

There's a psychological trap underneath it, too, the same one that makes slot machines profitable: intermittent rewards. Every so often a reroll lands something genuinely good, and that unpredictable payoff keeps you pulling the lever long after it's stopped being productive. Naming the trap is half the cure. The moment you catch yourself hitting generate 'just once more' hoping for luck, you've left the design process and entered a casino - and the way out is not another roll but a deliberate read of what you already have.

The alternative isn't more effort - it's more structure. Treat each generation as an experiment with one hypothesis, and the same hour produces a controlled understanding of exactly which words and parameters move your image which way. That understanding compounds; the pile of rerolls doesn't.

The method is small enough to hold in your head and strict enough to actually work.

1. Base prompt. Write one honest, slotted draft (Lesson 1.1) with high-signal words (Lesson 1.2). Don't polish it - you just need a starting point to react to. Generate it and, crucially, note the seed (Lesson 1.3).

2. Read the result. This is the step everyone skips and it's the most important. Look at what you actually got and name the single biggest gap between it and your intent - not five things, one. 'The light is too flat.' 'It reads as a museum, not a house.' 'The material looks like plastic.' You can't fix what you haven't named.

3. Change one variable. With the seed locked, make exactly one change aimed at that gap - swap one word, adjust one weight, change the lighting term. One. This is the rule the whole method rests on, because it's the only way the next step means anything.

4. Compare. Put the new result beside the old. Because only one thing changed and the seed held, any difference is caused by that change - so you've learned a fact: 'noon' flattens the shadows, travertine:1.4 makes the wall read. Keep the win or revert, then loop back to step 2 on the next-biggest gap.

The loop lives or dies on step 2, so it's worth its own beat. Reading an AI result is a real skill, and designers are trained for it - it's the same critical looking you do at a desk crit, pointed at a machine's output.

Discipline yourself to describe the gap in design language before touching the keyboard: composition, proportion, material read, light quality, mood, level of detail. Vague dissatisfaction ('it's not right') can't be acted on; a named gap ('the entry doesn't read as the primary threshold') points straight at the variable to change. Often the fix isn't adding words at all - it's changing an existing slot, removing a term that's fighting, or adjusting a parameter.

A useful habit: keep the intent visible. Write your one-line design goal at the top of your workspace ('a quiet, heavy, cave-like retreat - not a glossy villa') and judge every generation against that, not against 'is this a nice image'. Plenty of beautiful images are wrong for the brief, and the slot machine will happily hand them to you forever.

There's a subtler benefit to naming the gap out loud: it forces you to convert a feeling into a cause, and a cause points at a variable. 'It feels cold' is a feeling; 'the light is too blue and too even' is a cause; 'change the lighting term to warm raking light' is the variable. That chain - feeling to cause to variable - is the entire intellectual work of iteration, and it's a muscle designers already have from studio critique. Every lap you run trains it, and before long you read an AI result the way you read a plan: fast, specific, and already reaching for the one thing to fix.

Each controlled lap produces a fact; a log is where those facts accumulate into expertise. It needn't be elaborate - prompt, seed, the one change, and the result in a line ('seed 42 + noon -> flatter, colder; reverted'). Over a project this becomes a map of how your subject responds, and across projects it merges with your vocabulary document into a genuine personal instrument. The log is also what lets you reproduce a client-approved image weeks later, which rerolling never can.

Finally, method includes knowing when to stop. Iteration has diminishing returns, and past a point you're polishing pixels a client won't notice or, worse, chasing a fidelity these tools structurally can't give - a specific plan, a real structural logic, an exact existing building. When the remaining gap is structural rather than aesthetic, no amount of prompting will close it: that's the signal to switch tools, to the conditioning and control methods of Module 3. Prompting is for steering appearance; when you need to command geometry, you've reached the edge of this module - and the start of real control.

It's tempting to treat all this as a lab exercise, so here is how it lives in actual practice - say you're developing a concept image for a small retreat to show a client next week.

You start with a slotted base prompt and generate a small spread, picking the one that best matches your written intent, not the prettiest. You note its seed; that image is now your anchor. From there every session is laps of the loop: read the anchor, name the biggest gap ('the roof reads too commercial'), change one term, compare, keep or revert, log it. In an hour you've not only improved the image - you've built a short, honest record of why each decision was made, which is exactly the narrative a good crit or client conversation needs. 'We tested the roof three ways; the low monsoon-tile version held the domestic feeling best' is a designer talking, not a prompt-copier.

Two professional habits make this bulletproof. First, branch, don't overwrite - when a change works, save that image and its settings before the next lap, so you can always walk back a wrong turn. Second, separate exploration from refinement: early on, high chaos and wide rerolling are legitimate for finding a direction, but the moment you've found it, lock the seed and switch to disciplined one-variable laps. Confusing the two phases - rerolling when you should be refining - is the commonest way good work dissolves back into slot-machine noise. Know which phase you're in, and the method carries you from a blank box to a defensible, reproducible image the whole team can trust.