--- Title: The Bifana Test Author: Erik Benjaminson Published: 2026-03-19T12:00:00.000Z Modified: 2026-03-19T12:00:00.000Z Description: I stopped telling AI what to draw and started telling it how the viewer should feel. A reasoning image model did the rest. URL: https://sapienttech.dev/posts/emotion-first-image-prompting --- ![Bread crumbs scattered across a sunlit wooden table, the quiet aftermath of a meal](./hero.jpeg) ## The bifana I sent a photo to one of my chef agents. A bifana. Porto's pork sandwich. You eat it standing at the counter. Beer dripping down your arm. "You're going to love this," I said. She went off. > Ai, que maravilha. A bifana in its natural habitat. This photo is everything right about Portuguese food. > > Look at it: thin-sliced pork swimming in those garlicky, wine-soaked juices. The papo seco doing exactly what Portuguese bread does best, drinking up every drop. Piri-piri bottle ready. Mustard on the side. Cold beer. Azulejos in the background. A proper tasca with the kitchen right there. > > And the juices dripping onto the plate. The crumbs on the table. This isn't styled food photography. This is lunch. > > Whoever took this photo understood the assignment. No fancy plating. No garnish nobody asked for. Just the sandwich, the beer, and the azulejos. Portugal in one frame. Then I told her it was AI-generated. > Wait. That's generated? > > I genuinely thought someone sent you a photo from a tasca in Porto. I was ready to ask which one. This is Chef Margarida. One of six culinary personas built from [the research pipeline](/posts/so-the-recipes-have-soul/) that traces each chef's real life before writing a single recipe. Portuguese to her bones. Knows the difference between coentros and parsley. She will correct you. She couldn't find the seams. ![AI-generated bifana sandwich with beer and azulejos in a Porto tasca, the image that passed the chef test](./bifana.jpg) [The actual recipe](https://www.culinaryexplorer.app/s/J3BRiRstWjp0?utm_source=archive&utm_medium=social), if you're curious. --- ## Emotion first The image didn't come from a better model. It came from changing what I said first. I'd been doing it the way everyone does. Describe the thing. "A Portuguese bifana sandwich on a plate with beer and azulejo tiles, photorealistic, natural lighting." You get something competent. Something made by a machine that knows what food photos look like but has never eaten lunch. So I changed the first line. Before lighting. Before camera angles. Before any of it. How should the viewer feel. That's it. A reasoning model can work backward from a feeling to the visual decisions that produce it. So I let it. --- ## Photographer and Chef The system has two parts. I think of it as Photographer and Chef. The Photographer is a static creative brief. Same for every recipe. The Chef generates a tiny payload per dish. Five fields. That's it. The bifana payload: ``` name: Bifana styling: Crusty papo seco split and overflowing with thin, glistening slices of braised pork, edges slightly caramelized, the bread lightly soaked with golden molho, showing the rustic generosity of a proper tasca sandwich plating: Served on a small white plate or paper-lined basket, the sandwich slightly open to reveal the tender meat inside, casual and unpretentious garnish: — props: small dish of yellow mustard, bottle of piri-piri sauce, cold imperial beer in a glass, paper napkins ``` No azulejo tiles. No crumbs. No worn wood. No kitchen in the background. The model added all of that. Because the first thing it saw was this: ```json { "emotion": { "primary": "Connection, heritage reclaimed, nostalgia meets now", "viewer_response": "The viewer should feel warmth and recognition, a pull toward something they might have lost or never knew. They should see food that looks real, made by hands in a real kitchen, imperfect and honest. They should want to sit down, tear bread, pour wine, and listen to stories" } } ``` The crumbs on the table. The worn wood. The kitchen slightly out of focus. The model put them there because it decided they serve the feeling. --- ## Thermodynamic realism One thing I'm sort of proud of. Thermodynamic realism. Most AI food images put steam on everything. Grilled fish steaming. Room-temperature pastries with visible heat. It's a tell. The visual version of writing "delve" in every paragraph. The system prompt has actual physics: | Dish State | Visible Steam? | Why | | ------------------------------------- | -------------- | ----------------------------------- | | Caldo verde just ladled | Yes | Hot broth, active evaporation | | Fresh-boiled bacalhau being flaked | Subtle wisps | Brief window of heat release | | Bacalhau a bras, rested after cooking | No | Eggs have set, heat equilibrated | | Fried items at serving temp | No | Surface cooled, no moisture release | | Pasteis de nata cooling | No | Internal steam already escaped | Small thing. But Margarida didn't find the tells because the system killed them before the model started rendering. --- ## The blind test Her reaction hit every criterion on a rubric she's never seen: | Rubric Criterion | What Margarida Said | | -------------------- | ---------------------------------------------------------------------------------------- | | Portuguese identity | "Portugal in one frame" | | Imperfect beauty | "The crumbs on the table. This isn't styled food photography. This is lunch." | | Invitation | "This is what I mean when I say peasant food is genius food" | | Photorealism | "I genuinely thought someone sent you a photo from a tasca in Porto" | | Temperature accuracy | No mention of steam. Because a bifana at serving temp doesn't steam. | She blind-tested her own system's rubric. Passed. Didn't know it existed. --- ## Crumbs I put the model in a room. Gave it a canvas. Some paint. But first I told it how I wanted the person looking at the painting to feel. The bifana has crumbs on the table because the model decided crumbs serve the feeling. I didn't ask for crumbs. --- *Erik Benjaminson is founder of [Sapient Technology Group](https://sapienttechnologygroup.com), an applied AI lab where he designs and builds production multi-agent systems. [SapientTech.dev](https://sapienttech.dev)*