Taste Company

Models can reason. They cannot feel a room.

Every design decision is an emotional one wearing a technical costume. We are building the taste layer: the part of a model that decides what a thing should feel like before it decides what it looks like — and then knows where to look to get there.

See the taste layer run Our first product

Shipping today as claudetaste, a skill for Claude Code

Target feeling

01 — NAME IT

The emotion is the brief

Before layout, the model resolves the brief into an affective target and the adjectives that follow from it.

    02 — STUDY IT

    Humans already solved this

    It pulls work that provably produces that feeling in people, and reads what does the producing.

    03 — BUILD IT

    Emulate the mechanism

    Not the pixels — the moves. Same component, four feelings, four different objects.

    Evening Programme Doors at seven. The hall stays quiet until it doesn't. Reserve a seat

    The prompt never changed. Only the feeling did.


    Why emotions

    “Emotion is not a luxury on top of cognition. It is how cognition chooses.”

    Forty years of research, one inconvenient conclusion

    • Without affect, you cannot decide at all

      Damasio · Descartes' Error · 1994 — somatic marker hypothesis

      Patients with damage to the ventromedial prefrontal cortex kept their IQ and their logic, and lost the ability to choose. Given two acceptable options they deliberated forever. Taste is the same problem: a model with no felt preference has no ground to prefer one of a thousand valid layouts, so it returns the statistical average of all of them. That average is what everyone calls AI slop.

    • Design works on three levels, and two of them are feeling

      Norman · Emotional Design · 2004 — visceral, behavioural, reflective

      Norman's framing: the visceral level reacts before thought, the behavioural level covers use, the reflective level covers what owning the thing says about you. Models are trained hard on the behavioural middle — does it work, does it parse — and are effectively blind to the other two, which is exactly where taste lives.

    • What feels good is judged to work better

      Kurosu & Kashimura 1995 · Tractinsky et al. 2000 — the aesthetic–usability effect

      Across cultures, interfaces rated more beautiful were rated more usable, before anyone used them. Affect is not decoration applied after the engineering; it changes how the engineering is perceived and forgiven. Optimising a page for correctness while ignoring feeling optimises the wrong variable.

    • Positive affect broadens what you can see

      Isen et al. 1987 · Fredrickson 2001 — broaden-and-build

      Mild positive affect measurably widens associative range and creative problem-solving. Anxiety narrows it to the nearest safe answer. A model with no affective state defaults to permanently narrow: the purple gradient, the centred serif, the three-card row. Giving it a feeling to hold is what widens the search.

    • People judge credibility on look, in seconds

      Fogg et al. 2003 · Lindgaard et al. 2006 — 50ms to a verdict

      In a study of over 2,500 people, visual design was the most cited reason for judging a site credible — ahead of the content. Separate work put reliable aesthetic verdicts at 50 milliseconds. That verdict is emotional, it is instant, and it happens long before your copy is read.

    The method

    Feeling first. Evidence second. Pixels last.

    A model does not need to have emotions to use them. It needs to treat the intended emotion as the primary constraint — and then reason from the human work that already achieves it.

    01

    Determine the feeling

    What should someone feel in the first second, and what should be left behind after they leave? Grounded in the subject's own world, not a generic mood board. A coffee roaster gets burlap and steam. A hospice gets stillness.

    02

    Learn from what already works

    Humans have spent centuries producing awe, urgency, tenderness and trust on demand — in posters, buildings, films, book covers. We index that work by the feeling it produces and read the mechanism behind it.

    03

    Emulate the mechanism, not the surface

    Copying a reference gives you a pastiche. Copying why it works — the pacing, the restraint, the single loud element — gives you something new that lands the same way. That transfer is the taste layer.

    Shipping

    The layer, in pieces you can install.

    claudetaste

    Available now

    A skill for Claude Code that runs the whole method before it writes a line of CSS: pick the feeling, refuse the defaults, spend boldness once. Same model, same prompt, work you'd actually ship.

    See claudetaste

    The rest of the layer

    In the workshop

    Design is the first surface, not the last. The same loop — name the feeling, study the humans who produce it, emulate the mechanism — applies to prose, product, sound and everything else a model currently averages its way through.

    Talk to us