Harvey

45 min · two projects

Open

Before they read anything

“Both of these stop at the system rather than running end to end. What they do show is the rules running as code.”

Lumasis — show one thing

Derived tokens

“Naming is where design systems rot. If names are assigned, two people name the same color differently and it forks. So the name is a function of the value — hue family plus lightness step. Two designers land on the same token independently.”

Standards Manual — show one thing

The coverage table

“This isn’t a picture of a table. Layout is one custom property redefined at four breakpoints, and every cell names itself when the header drops. I added a Hebrew row last week — one edit, and the stagger, the totals and the CSV all followed.”

“That’s my test for whether I built a component or drew a picture of one.”

The one answer you can’t improvise

If they ask how you use AI

“I specify the constraint, the model executes, and I verify against the live page rather than what it reports back. I moved a control-room map onto new tokens — I didn’t pick colors, I described the semantic mapping, then checked it: thirty-three bindings, no semantic token holding a raw value. It only works if the verification is real.”

Ask them

Who consumes the system — components, tokens, or docs?
When a team needs something the system doesn’t have, what happens?
Where does generation sit in your workflow in a year?

If asked about adoption: not on these pages. A fabricator, a wrap vendor, one engineer — not designers pulling instances. It’s what you’d build first here.

Never cut your questions.