ONE · WHAT WENT OFF TODAY Google's documentation puts it in one sentence: the Imagen 4 standard, ultra and fast endpoints are deprecated and will be shut down on 17 August 2026, and you should move to Gemini 3.1 Flash Image to avoid interruption. Three named things stopped today, and if you built anything against them by name, it stopped with them. That much is an ordinary retirement, and if it were only that it would be a footnote. The interesting part is what the replacement is not. Gemini 3.1 Flash Image is the model most people know as Nano Banana 2, and it does not have a separate image request. The old dedicated method for making pictures is gone. Images now come back from the same request you use for words. Same door, different thing carried through it.  GO TO THE SOURCE The retirement notice and the model list are here. What the request will actually accept is documented separately, here, and the second page is the one worth your ten minutes. Read the input section rather than the model comparison. The comparison tells you which one is quicker. The input section tells you what you are now allowed to hand it, and that is the part that changes what you can make. TWO · WHAT THE BOX TAKES NOW Four things, and only the first is the one everybody uses. Words, as always. Up to fourteen reference images in a single request. A video, which the model reads for context, for who is in it and what they are doing, and turns into stills. And a link to a video rather than the file itself. Fourteen is the number to sit with. For two years the working assumption in image prompting has been that you describe in words what you cannot show, because you were allowed one reference if you were lucky. Fourteen is not one more. Fourteen is enough to stop describing and start showing: the face, the garment, the room, the light, the lens character, the palette, the thing on the shelf in the corner that you kept failing to get across. And a video reference is a different kind of instruction again, because a video carries motion and continuity that no still can. Handing over three seconds of someone walking says more about how they hold themselves than a paragraph ever will. THREE · HOW TO WRITE FOR A REFERENCE YOU DID NOT WRITE Here is the habit that will hurt you. Most of us learned to write prompts that carry the whole load, because nothing else was carrying any of it. Paste that same paragraph in alongside fourteen references and a video, and you are now issuing two sets of instructions that disagree, and the model has to pick. So the discipline reverses. With references attached, the words are not the description any more. They are the difference. Say what the reference does not already say, and say what you want changed about it. Everything else you type is noise competing with a picture, and the picture is more specific than you are. The practical version is a subtraction pass. Write the prompt you would have written, then delete every line the references already answer. What survives is usually three sentences, and it will hold up where the paragraph did not. PROMPT OF THE DAY The one that says only what the reference does not. Reach for this the first time you attach several references or a clip and notice your usual paragraph fighting them. Attach the references, then paste this with your old prompt underneath it. I am giving you reference material and a prompt I wrote before I had it. Do not generate anything yet. First, list what the references already establish on their own: subject, clothing, setting, light, palette, camera distance, and anything about how the subject carries themselves. Second, go through my prompt line by line and mark each line as already covered by the references, in conflict with them, or genuinely new. Third, rewrite my prompt using only the lines you marked genuinely new, plus any change I am asking for that the references do not show, and keep it under four sentences. Tell me plainly which of my lines you dropped and which one conflicted, because the conflict is the thing I most need to know about. The load-bearing instruction is the conflict flag. Covered lines are merely wasted, and waste costs you little. A line that argues with the reference is the one that quietly produces the result you cannot explain, and it is almost always a habit you carried over from when words were all you had. THE NEXT WORKING NOTE Every capability that arrives quietly changes a habit that arrived loudly. Long descriptive prompting was never a craft principle. It was a workaround for not being able to show the model what you meant, and we mistook it for skill because it was hard and we got good at it. When showing becomes possible, most of that skill turns into overhead, and the people who keep doing it will wonder why their results got less predictable rather than more. I am putting together a five case input sheet: the same shot attempted with words only, with one reference, with several, and with a clip, so you can see exactly which sentences stop being necessary at each step, each with a worked example. Built so you can tell in advance which half of your prompt to delete. Want it when it ships? Reply with send me the input sheet and I will get it to you. A QUESTION FOR YOU How long is your usual image prompt, honestly. A line, a paragraph, or something you have been quietly refining for a year. I am asking because the answer decides how much of this issue actually applies to you, and I suspect the people with the most refined paragraphs have the most to unlearn. | 