how-i-build
A gardener in a dark blue apron pruning a cloud-shaped pine by hand, over a bed of moss, in a Japanese garden with a wooden house behind
Design

One skill to choose, another to prune

I paired two design skills in Claude Code: one sets the direction, the other builds and reviews. Each is good on its own. Together they gave me, in minutes, a site I would not be embarrassed to show.

4 min read


Design has always been the part of a project where I lose the most time and trust the result the least. I know when a page is ugly. The hard part is knowing what to do so it stops being ugly.

Over the last two days I used two design skills in Claude Code, together, on two projects. I came out more satisfied than with anything I had tried for this before, Claude Design included. This post is about why the pair works better than either one alone.

What each one does

Taste is an anti-generic skill. Before writing a line, it reads the request: what kind of page it is, who it is for, which words I used, what brand already exists. Then it picks a direction and defends it. Its whole point is not to fall into the default look every AI-made site ends up with.

Impeccable is an execution skill with a high bar. It builds the page, takes screenshots on desktop and mobile, runs a defect detector and sends everything to a reviewer with a clean context, which returns an ordered list of fixes. Once the page passes, another agent writes the DESIGN.md from what shipped, not from what was meant to ship.

Both are good alone. The problem is that both have opinions.

Why together they need a rule

If both skills propose a direction, they fight. Taste suggests one thing, Impeccable reinterprets it, and the page becomes a compromise nobody chose.

The rule that fixed this was splitting the work:

  • Taste decides. It locks the visual direction: type, color, what the page's protagonist is.
  • Impeccable builds and checks. It does not reopen the direction. It builds on what was decided and judges the page against that decision.

With that split, each one does what it does best. One has taste, the other has rigor. And the review finally has something to measure against: the reviewer does not ask "is this pretty?", it asks "does this hold the direction that was locked?". That is a question you can answer.

The direction I would not have chosen

The best moment was when Taste disagreed with me.

I wanted a landing page for LyricsLens, an app that shows a song's lyrics over the screen. I came in with the idea of a 3D logo spinning in the center. Taste read the request and proposed something else: the protagonist would be a line of the lyrics, sung karaoke-style, and the 3D logo would be only a detail.

It was right. The page is about lyrics, so the lyrics had to be at the center. The logo ended up small, in the corner, the size of an icon, and you can spin it with the mouse. I would never have got there on my own, because I was in love with the wrong idea.

Counted rounds, not a loop

The other thing that surprised me was Impeccable knowing when to stop.

It works in closed rounds: build everything, inspect once, fix everything in one batch, confirm at most once more, and stop. No polishing forever. And the reviewer is honest about what it looked at. In one round it approved six fixes and wrote down that the approval covered only those six, not the whole page. I prefer that to a generic "all good".

What the reviewer caught were things I would have let through: a blue that read as GitHub's default blue instead of the cobalt that was chosen, a button drawing more attention than the main information, a table that clipped values on mobile.

The fast part

On keepalive, my service monitor, I asked for a new design from scratch. From the first message to the page approved by the reviewer took about fifteen minutes, running on localhost against real production data. Twenty minutes after the first message, it was live.

On the LyricsLens landing page, the reviewed page was ready in a little over half an hour. The rest of the afternoon was me adjusting what was my own taste: the karaoke filling word by word, the line below rising to become the main one, looping videos instead of screenshots, the small white treble clef. The skills did not take those decisions away from me. They took away the work of getting to them.

Why this was better for me

What comes out of these two skills is not a drawing I translate later. It is the site itself, in the repository, running on my machine, with real data. Every change I ask for shows up in the browser right away. And at the end there is a DESIGN.md describing the visual system that shipped, so the next change does not start from zero.

The gardener in the photo does not decide the tree's shape while he cuts. The shape was chosen before, and every cut serves it. That is what I got by splitting the two skills: one chooses the shape, the other prunes until it shows.