Every morning I make the same cup of coffee. Same dripper, same kettle, same scale. And almost every morning it tastes a little different.
For a long time this annoyed me. Now it is my favourite part of the day.
Four knobs, one cup
A pour-over has very few moving parts. You have the beans. You have the grind size. You have the water, its temperature and how much of it. And you have time.
That is it. Four knobs.
But those four knobs interact in ways that are hard to predict. Grind a little finer and the water moves slower, so the brew takes longer, so more comes out of the coffee. Pour a little faster and the opposite happens. Change the beans and everything you learned yesterday is only half true today.
If the cup is sour, it probably ran too fast. If it is bitter and flat, it probably ran too slow. You adjust one thing, brew again, and taste.
This is exactly how I work with AI
A few months ago I noticed that I talk to coding agents the same way I make coffee.
A prompt also has a few knobs. The context you give. The instructions. The examples. The constraints. Change one, and the result shifts in a way you did not fully expect.
When the output is too vague, I add context. When it is too clever and goes off in its own direction, I add constraints. When it keeps missing the style I want, I show an example. Then I run it again and look at what came out.
Nobody gets the perfect prompt on the first try. Nobody gets the perfect cup on the first try either. The good ones are not written. They are tuned.
Change one thing at a time
The single best coffee lesson I ever got was boring: change one variable at a time.
If you grind finer, use hotter water, and switch beans all on the same morning, the cup will be different, and you will have no idea why. You learned nothing. You just rolled the dice.
The same is true for prompts. When something does not work, the temptation is to rewrite the whole thing. New wording, new structure, new examples, all at once. Sometimes it gets better. But you do not know which change mattered, so next time you start from zero again.
Small, single changes feel slow. They are actually the fastest way to understand what you are doing.
Taste is the real skill
Here is the part that took me longest to accept. The scale and the timer do not make a good cup. They only make the process repeatable. The real skill is tasting the result and knowing what is wrong with it.
Sour or bitter. Thin or heavy. Bright or muddy.
With AI it is the same. Anyone can type a request. The value is in reading the answer and knowing, quickly and honestly, whether it is good. Is this code actually correct, or does it just look correct? Is this text saying something, or just sounding confident?
If you cannot taste the difference, no amount of tuning will help. You will keep pouring and never know if you are getting closer.
Write down what worked
Coffee people love their notes. Bean, grind setting, ratio, time, and one line about how it tasted. It looks a bit obsessive until you realise that after two weeks you have a recipe that works almost every time.
I started doing the same with prompts. Not a fancy library. Just a file with the ones that worked, and a short note on why. It is the closest thing I have to a recipe book for my work.
The morning ritual
So every morning I weigh the beans, grind them, heat the water, and pour slowly. I taste it and think: a little finer tomorrow.
Then I open my laptop and do the same thing with words.
Neither process is magic. Both reward patience, small changes, and an honest tongue. And both get a lot better once you stop expecting the first try to be perfect.
