Prompt-writing looks like a vocabulary problem — collect enough magic words and you're set. In practice it's an observation skill first: most weak prompts are weak because the writer didn't notice what the image in their head actually contains. Here is a four-step practice path that builds that noticing, using only what's on this site.
Step one: name what you see
Promptle is the easiest place to start, and the training is in the misses. Play a few rounds and watch which words you guessed: if your guesses keep being captions ("a girl near water") and the keywords keep being physical things ("pier", "umbrella", "cap"), you've found exactly the habit to break. The how-to piece covers the three habits that speed this up, but the game itself is the exercise.
Step two: read the notes
Every searchable prompt page has an editorial note under the image. Read five of them in a row, in categories you don't usually look at. The point is not the specific prompts — it's watching the note split an image into subject, styling, camera, mood and format, over and over, until you start doing it automatically. The five-moves teardown explains the underlying structure the notes follow.
Step three: change one layer
Pick one prompt from the gallery and copy it exactly. Generate once. Then change exactly one layer — swap the wardrobe, or the light, or the mood word — and generate again. Put the two results side by side. This is the fastest way to learn what each layer is doing, because you hear each one speak alone. Changing everything at once is how most people end up with prompts that work by accident and can't be fixed on purpose.
Step four: write from your own desk
Now write one from scratch, about an object you can actually see — a mug, a plant, your shoes. Follow the brief in order: subject first, then what the background does, then where the light comes from, then the mood, then the format. Don't reach for the mood words until the layout is settled. Compare with what the gallery would have said. The full blank you're filling in is described in the corpus audit — the vocabulary the library actually uses, in the order it uses it.
None of this needs a special tool. One daily round of Promptle, five notes read closely, two generations per prompt — that's the whole loop, and it compounds faster than collecting word lists ever will.