Coming soon

Salt to Taste

How to teach an AI your eye for design taste, without touching a single model weight.

The best bread recipe I own was handed down through three generations, and the last line of it is a lie. Not the ingredients. Those are honest, measured to the gram. The lie is the instruction at the end. Salt to taste.

Which is not an instruction. It is a bill sent to a palate the reader does not have yet. The woman who wrote it had salted ten thousand loaves and her hand knew the amount, so three words were enough to mean a number. To anyone else they are noise.

That recipe is every prompt you have ever given an AI to design an interface. You type make it clean and modern, and the machine hands back something flat: technically a UI, structurally sound, completely forgettable. Then you spend the next hour repainting it by hand and tell yourself the AI is not there yet. The AI is there. You sent it a bill for a palate it does not have.

This is a short book about paying that bill. It takes the taste that lives in your hand, the thing you can see instantly in someone else's design but cannot quite explain, and writes it down in a form an AI model can actually obey. Nothing here touches the model. You train the context instead, in four moves.

  1. Tell Write the rules down with numbers in them, so there is nothing left for the model to resolve toward the middle.
  2. Show Give it the anchors to reason toward, and the refusal list that names what you will not accept.
  3. Test Judge the work with something that did not write it, starting with a check mechanical enough to run in CI.
  4. Keep Every correction becomes a rule, so you never have to give the same note twice.

The appendix is the prompts on this site, which is why they are already here in full rather than waiting on a publication date.

Buy the book

Nine chapters, the appendix and the footnotes, in one file that reads on a phone. Everything on this site stays free either way.

Coming soon
The chapters

Contents

Nine chapters, two openings, one appendix.

The four moves are chapters 4 to 7 and they build on each other, but nothing here has to be read in order. The first three chapters are the argument for why the fourth is worth the afternoon.

  1. i Front matter About this book Who the book is for, and the fair warning that comes with it: you are about to find out how much of your taste was a vibe.
  2. ii Front matter Salt to taste The bread recipe whose last line is a lie, and why that lie is every prompt you have ever written.
  3. 1 Chapter 1 The fastest intern on earth The moment the machine builds a real page in two minutes, and the second and third pages that turn out to have the same bones.
  4. 2 Chapter 2 Why your AI has no taste A model is drawn to the center of what it has seen, and one sentence asking for taste, exploration, and code at once points it straight there.
  5. 3 Chapter 3 Taste is teachable and articulable If you could brief a junior designer well enough that they improved, your taste is already articulable, and a file is only a junior who never forgets.
  6. 4 Chapter 4 Write rules with numbers in them The difference between a rule that trains a machine and a rule that trains nothing is almost always a number.
  7. 5 Chapter 5 Show. The banned list beats the mood board Positive examples set the aim. The explicit nos do the heavy lifting, because rejection is definition.
  8. 6 Chapter 6 Test. Never let it grade its own homework Three layers of judgment, cheapest first, and not one of them written by the thing that made the work.
  9. 7 Chapter 7 Keep. Taste that compounds The small loop that turns each correction into a written rule, which is how salt to taste finally gets a number next to it.
  10. 8 Chapter 8 Five sessions, unedited Five real builds with the names filed off and the specifics kept, including the ones that went wrong first.
  11. 9 Chapter 9 Steel-man, the pitfalls, and Monday The strongest argument against the whole method, taken seriously, and what to do on Monday morning anyway.
  12. A Appendix Steal everything Six starter artifacts, worth more edited than obeyed. Every one of them is on the prompts page, kept current as the models change.

Colophon. The prompts on this site and the appendix in the printed book are generated from one set of files in the Workshopr repository, so the page you are reading and the page you are holding say the same thing. Source last changed 24 September 2026.