How to Beat a Single Prompt

# A 12B model can see your food. It cannot count the calories.

### So I stopped one-shotting it. The same model, wrapped in a small workflow, goes from useless to within 14 calories of the truth. A measured case for workflows over raw prompting on local models.

I wanted to compare two ways of guesstimating the calories in a photo with a small
local model: one-shotting it, versus breaking it into a multi-step workflow. I
scored it against Nutrition5k, a dataset of real plated meals whose calories were
measured in a lab. I chose to run this locally on a small 16GB 4080 in lieu of using
Claude or ChatGPT to better determine gains.

## The OneShot way: ask the model

Show gemma4:12b a photo and ask for the total. It answers instantly and the estimate is not close. Across 24 dishes its mean error was
504 calories, and it tends to guess somewhere between 700 and 1,250 almost
regardless of what is on the plate.

The model is not blind. Ask it to describe the same photo and it is near perfect:
"roasted fish, potato wedges, carrots, a side salad." It sees the food fine. It
just cannot turn what it sees into a calorie number, because that number is a fact
it half-remembers, not something it can read off the image.

## The Multi-step way: a workflow

The workflow splits the
job along that exact seam.