One prompt is not a process - what separates AI slop from work you can ship

What slop looks like from the inside

Everyone recognizes slop from the outside: the waxy skin, the gradient soup, the composition that feels like it was assembled by committee-of-none. From the inside, slop has a simpler definition. It is output that went straight from prompt to publish.

The tools invite exactly that. Every generator demo shows a one-line prompt becoming a polished frame in seconds, and marketing teams under deadline pressure take the demo at its word. The result floods feeds, and serious clients now flinch at "AI-generated" because the label covers both a throwaway post and a broadcast-grade spot, with no way to tell from a thumbnail which one you are about to get.

The fix is not a better prompt. It is admitting that the demo was never the process.

What one finished image really takes

Agencies that publish their kitchens - Level Agency did this recently with a mock product shot, and their write-up matches our experience shot for shot - land on the same shape: something like eight distinct steps for a single finished image. Not eight prompts. Eight stages, each with a human decision in it.

Ours looks like this on a typical product visual. Moodboarding in an image model to find the visual territory. A dedicated pass to get the product itself right, because generic models mangle labels and proportions. Selection - the step where two hundred candidates become three. Compositing the approved product into the approved scene. A realism pass. Upscaling. Retouch and finishing in a proper design tool, by a designer, on layers. And a final review against the brief before anything leaves the building.

Strip any one of those and you can see it in the file. Strip four and you have slop with good lighting.

We don't generate. We produce. The sentence is on our site because it compresses everything above into five words.

The questions we ask before anything ships

Process only helps if there is a gate at the end. Before an AI-assisted asset goes to a client, someone here has to answer for it:

• Would this hold up printed large, not just at feed size?

• Is the product true - label, proportions, materials - or merely plausible?

• Does it sit inside the brand's world, or inside the model's default aesthetic?

• Are hands, text, edges and reflections clean at 200% zoom?

• Could a competitor have generated this exact frame? If yes, it is not distinctive enough to ship.

Five questions, thirty seconds each. They kill perhaps a third of work that looked fine at first glance, and that third is precisely the gap between us and a prompt.

Curation is the job

Here is the uncomfortable part for anyone selling "AI at the push of a button": the model contributes volume, and volume has almost no value. Value concentrates in the choices - which direction, which candidate, which crop, which imperfection to keep because it reads as human.

That choosing cannot be delegated to the tool that produced the options. It needs someone with references in their head and a stake in the outcome. Which is why the honest pitch for AI creative is not "faster and cheaper with no people". It is "faster and cheaper because the people moved to where they matter - the brief and the edit".