ONE · WHAT WAS ACTUALLY PUBLISHED
Seven briefs, run through Recraft V4.1 and through Ideogram V4, with both results shown side by side and a short verdict under each pair. The subjects are a kitchen at full service, a flat vector gag about a pizza box at airport security, a skincare product still life, a stylised caricature portrait, a young witch mid-spell, and a poster where type and illustration have to interlock.
Recraft wrote it, Recraft ran it, and Recraft decided who won. Every verdict favours Recraft. That is not a scandal, it is simply what a company blog is, and the honest response is to take the verdicts as marketing and keep reading, because the briefs underneath them are unusually good and they are printed in full.
Whoever wrote those prompts knows the work. They are not tag soup. Each one carries a subject, a camera position, a light source with a direction, a named style, and a constraint that can actually be checked afterwards. That last part is what makes them useful to borrow.
GO TO THE SOURCE
The seven briefs and both sets of results are here. Read the prompts, look at the pairs, and form your own view before reading the caption underneath each one.
TWO · WHAT THE BRIEFS GIVE AWAY
The caricature brief is the one to study. It runs to well over two hundred words and it never once reaches for a taste word. No striking, no beautiful, no masterpiece. Instead: three-quarter profile facing right, pale cool-blue skin with grainy crayon texture, a thick blocky red nose with defined nostrils, eyes looking right and downward, a coral-red clenched hand supporting the chin with three short knuckle lines, lighting stylised from the top left with no realistic shadows.
Every one of those is checkable. You can look at the returned image and say yes or no. That is the whole difference between a brief and a wish, and it is why a long prompt written that way keeps working while a long prompt made of adjectives falls apart.
The witch brief does the same thing in a different register. Very long arms, very short torso is a proportion instruction, and the write-up notes that one model followed it and one did not. That is a real observation precisely because the instruction was specific enough to fail.
THREE · HOW TO READ ANY VENDOR TEST
Three questions, and this post answers them the way most do. Who chose the briefs. Who ran the other model, and on what settings. Who wrote the caption. When the same party answers all three, the result tells you what that party wants shown, and nothing about which tool suits your work.
There is one claim here that does survive, because it is structural rather than aesthetic. The vector brief turns on whether the output is a real editable path or a picture of one, and that is a file property, not a matter of taste. A verdict you can check by opening the file is worth more than six verdicts about which kitchen felt busier.
So the sorting rule is simple. Keep the claims you could verify yourself in under a minute. Set aside the ones that need you to share somebody's taste.
PROMPT OF THE DAY
The one that turns your own work into a test set.
A vendor's briefs are chosen to suit the vendor. Yours are chosen by your clients, which makes them the only comparison that answers your question. Paste three of your own recent prompts in after the instruction.
Here are three prompts from my recent work. Turn each one into a test I can score without arguing with myself. For each prompt, list every instruction that is checkable by looking at the returned image, phrased as a yes or no question: subject present, camera position, direction of the light, count of objects, proportion, palette, whether text is legible. Then list every instruction that is a taste word and cannot be scored that way. Return two columns per prompt and nothing else. Do not improve my prompts. Do not tell me which model to use.
Run the yes or no column against two models on the same day, count the yeses, and you have a comparison built from your own work rather than somebody's launch post. The taste column is still worth having, but it belongs to you and it does not belong in a score.
THE NEXT WORKING NOTE
A launch post is an argument, and the prompts inside it are evidence the author had to show in order to make the argument. That is why the prompts are usually the honest part of any comparison, whoever published it. Collect them. Ignore the captions.
I am building a brief bank: the seven briefs from this post plus the ones worth keeping from every model comparison published this year, each stripped of vendor framing and marked with the single checkable constraint it turns on. Reply with send me the brief bank and I will get it to you when it is done.
A QUESTION FOR YOU
When you last switched image models, what actually decided it. A published comparison, a friend, or something you ran yourself. I am asking because I suspect the honest answer is mostly the second one, and that is worth knowing about all of us.
| 
