Interface

Build it by loading, not by inventing.

The problem this exists for, in Caleb's words: "right now it's laggy because it's needed to every time recalibrate how to make something." The cost of generating a UI is not rendering. It is the model re-deciding, per request, what a dashboard is, what spacing to use, what a good table looks like and where the data comes from. All four of those are written down here, once.

The order, every time

1. Name the template. references/taxonomy.md has ten. The request is almost always one of them, and the one most people mean is Dashboard.
2. Name the organisms it holds, from the same file. Stop here if you are about to invent a component that already has a name.
3. Find the data in references/data-loaders.md, and when it is not there use references/discover-data.md to go find it, appending what you learn back into the loaders file so the next session does not repeat the search.
4. Apply the system, from crafts/interface.md: grayscale first, the constrained scale, hierarchy by size and weight and color, too much whitespace then remove, details last.
5. Build the three states. Loading, empty, error. This is the step that gets skipped, because the prompt only ever describes the happy path.

Steps 1 to 3 are lookups and should cost almost nothing. That is the point.

Pick the stance from the reader, not from taste

On 2026-09-27 Amber Circles passed ux-lint, axe, layout-check and design-gate in the editorial stance (serif display, one gold italic phrase, an uppercase kicker over every heading, outline buttons) and Caleb called it AI slop. It was: that look is now the median "tasteful" generated page, and every gate scores the absence of the old tells, not this one.

The fix that landed was choosing the stance from who reads it. The users were older people opening links from texts, so it became signage (ux-engine material) in Atkinson Hyperlegible, the Braille Institute's face for low vision. Two rules came out of it:

• Give the accent exactly one meaning ("you can tap this") and write a test that fails when it fills anything else. The first signage landing had four amber things above the fold and design-gate read it LOUD at ink 0.52.
• Show the content as itself. The text a group actually receives, each tool as a live thumbnail the way Docs shows pages, faces instead of counts. A description of the product is the generated move.

One theme, never re-chosen

Caleb: "keep it the same theme." Tokens are defined once per surface and referenced everywhere. A generator that picks a fresh accent color per artifact produces a pile of demos rather than a product, and the tell is obvious the moment two of them sit side by side.

Where the thing renders

Pick the surface before the layout. An overlay HUD and a settings page are different templates, and building the second one into the first is the common mistake.

Reference files

• references/taxonomy.md: every component and template, named, grouped by the job it does. Read at steps 1 and 2.
• references/data-loaders.md: every data source already reachable on this machine, and the four shapes any new one arrives in. Read at step 3.
• references/discover-data.md: the procedure for a source nothing covers yet, cheapest method first, ending in pixels only as a last resort. Read when step 3 misses.
• crafts/interface.md: the researched rules, from Refactoring UI and Atomic Design. Read at step 4.

The hard line

Never invent a component that has a name in the taxonomy. A bespoke variant of a table is the recalibration this skill exists to prevent, it will not match the next one generated, and nobody asked for it.