ui-variant-lab
Produces N genuinely different landing-page (or brand-surface) variants from one
frozen content contract, serves them side by side, and hands the choice back to
the user. The deliverable is a decision, not code — until a direction is picked,
nothing touches production.
Use this at the direction stage. Once a direction is chosen and the question
narrows to "which version of this component," the problem is different: vary one
axis at a time (structure, density, emphasis, type, voice) so the result is
attributable. Jakub Krehel's variant skill owns that later stage.
The contract comes first, briefs second
Before any variant exists, write one REQUIREMENTS.md that pins:
- Verbatim copy: nav labels, hero H1/sub, section kickers and H2s, pricing
numbers, FAQ answers, CTA text and destinations. Variants may not paraphrase.
- Required sections and their order.
- Banned words and banned claims (no invented customers, counts, logos,
ratings, or benchmarks).
- The floor: real interactive elements, focus-visible,
aria-expanded on
accordions, prefers-reduced-motion end states, no-JS readable, primary CTA
above the fold, breakpoints at desktop/tablet/phone widths.
- Tokenization: all color/space/type through
:root custom properties so
directions stay swappable.
Then one brief per direction. A direction names a world — palette register,
type pairing, density, shape grammar, and 2-3 signature devices — not a recolor.
"Dark swiss" and "dark swiss but teal" is one world; "dark swiss" and "warm
paper notebook" are two. If two briefs would produce indistinguishable pages,
merge them.
Named worlds sourced from real references beat invented ones. Say which
reference a brief draws from so the choice is honest.
Build in parallel, in isolation
- One self-contained
.html per direction, built by separate workers
(subagents, worker CLIs) against the contract + brief only. Builders must not
see sibling variants — convergence kills the point.
- Variants never import production code, and production never imports them.
- Serve the directory over HTTP (
python3 -m http.server, bunx serve) and
ship an index.html gallery that opens each variant full-screen. Thumbnails
distort spacing; judge at real size.
Review with screenshots, then stop
Capture the same matrix for every variant: hero at ~1440×900, a full-page pass,
and 390px. Watch for content clipped at the right edge — overflow-x:hidden
hides the scroll but not the lost glyphs; the check is whether words are cut,
not whether the page scrolls.
Then present a tradeoff table and stop:
| Variant |
World |
Right when |
Costs |
| Signal |
Dark console, live feed |
The differentiator must be visible |
Reads operational, not warm |
| Notebook |
Warm paper, annotations |
Trust through scholarship |
Quietest, least product feel |
Never mark a favorite in the table. Asked directly, answer from the product's
differentiator and audience, not which file was most fun to build.
Promote, then delete
On a choice: port the winner into the real stack (components, tokens, routes),
keeping the contract's copy and the floor's checks. Delete the losing files and
the gallery unless the user asks to keep them — harness files that linger get
mistaken for product.
Outside references
The project's own design system decides first. These fill the gaps it has not
decided, most often when picking direction worlds. Fetch the real content
before citing a reference — if a page cannot be read, say so and design from
the system's own tokens. Never imitate from memory. Report which reference was
used and what it changed.
Style directions:
| Reference |
What it gives |
| styles.refero.design |
2,000+ named DESIGN.md style directions from real product sites — colors, type, spacing per world. The fastest way to name a direction. |
| VoltAgent/awesome-design-md |
Open-source DESIGN.md collection for ~74 brands, MIT. |
| minimal.gallery |
Curated minimal-site gallery; tag pages (/tag/saas/, /tag/startup/, /tag/one-page/) for restraint baselines. |
Components and interaction (full list in ui-micro-interactions):
| Reference |
What it gives |
| 21st.dev |
React + Tailwind component registry with an MCP; free usage limit — say what you are looking for before calling. |
| component.gallery |
60 components across ~95 design systems. Check how mature systems handle a component before inventing one. |
| microkit.co |
MIT micro-interactions on the shadcn registry. |
| aicss.dev |
Agent-UI component taxonomy (thinking, tool calls, streaming, approvals). Pro source — use the taxonomy, not the code. |
Motion and video:
| Reference |
What it gives |
| kinetics.colorion.co |
Spring-physics motion recipes; copy CSS, React or an AI prompt. |
| whatships.com |
Launch-video library; tile a reference's frames to study pacing. |
| hyperframes.dev |
Renders an authored HTML sequence to MP4 for demo videos. |
| impeccable.style |
polish/distill/bolder review commands for "still feels off" work. |
| jakubkrehel/skills |
better-* interface skill collection — npx skills add jakubkrehel/skills. |
Before you finish
| Mistake |
Fix |
| Variants differ only in accent or font |
Merge them; write a different world |
| Builders saw each other's output |
Rebuild the converged ones blind |
| Judged from thumbnails or file:// |
Serve over HTTP, review at real size |
| A variant clips words at 390px |
Fix the floor before it enters the gallery |
| Table marks a favorite |
State each world's cost; the user picks |
| Harness survives promotion |
Delete it in the promotion commit |
1---2name: ui-variant-lab3description: "Explore a landing page or brand-level redesign by building several whole visual worlds against one frozen copy contract, comparing them in a local gallery, then promoting the winner. Use when a page needs a new direction, the user cannot name a style, or several candidates must be compared side by side."4metadata:5 source: https://skillmd.com/skills/pooriaarab/ui-variant-lab6---78# ui-variant-lab910Produces N genuinely different landing-page (or brand-surface) variants from one11frozen content contract, serves them side by side, and hands the choice back to12the user. The deliverable is a decision, not code — until a direction is picked,13nothing touches production.1415Use this at the *direction* stage. Once a direction is chosen and the question16narrows to "which version of this component," the problem is different: vary one17axis at a time (structure, density, emphasis, type, voice) so the result is18attributable. Jakub Krehel's `variant` skill owns that later stage.1920## The contract comes first, briefs second2122Before any variant exists, write one `REQUIREMENTS.md` that pins:2324- Verbatim copy: nav labels, hero H1/sub, section kickers and H2s, pricing25 numbers, FAQ answers, CTA text and destinations. Variants may not paraphrase.26- Required sections and their order.27- Banned words and banned claims (no invented customers, counts, logos,28 ratings, or benchmarks).29- The floor: real interactive elements, focus-visible, `aria-expanded` on30 accordions, `prefers-reduced-motion` end states, no-JS readable, primary CTA31 above the fold, breakpoints at desktop/tablet/phone widths.32- Tokenization: all color/space/type through `:root` custom properties so33 directions stay swappable.3435Then one brief per direction. A direction names a *world* — palette register,36type pairing, density, shape grammar, and 2-3 signature devices — not a recolor.37"Dark swiss" and "dark swiss but teal" is one world; "dark swiss" and "warm38paper notebook" are two. If two briefs would produce indistinguishable pages,39merge them.4041Named worlds sourced from real references beat invented ones. Say which42reference a brief draws from so the choice is honest.4344## Build in parallel, in isolation4546- One self-contained `.html` per direction, built by separate workers47 (subagents, worker CLIs) against the contract + brief only. Builders must not48 see sibling variants — convergence kills the point.49- Variants never import production code, and production never imports them.50- Serve the directory over HTTP (`python3 -m http.server`, `bunx serve`) and51 ship an `index.html` gallery that opens each variant full-screen. Thumbnails52 distort spacing; judge at real size.5354## Review with screenshots, then stop5556Capture the same matrix for every variant: hero at ~1440×900, a full-page pass,57and 390px. Watch for content clipped at the right edge — `overflow-x:hidden`58hides the scroll but not the lost glyphs; the check is whether words are cut,59not whether the page scrolls.6061Then present a tradeoff table and stop:6263| Variant | World | Right when | Costs |64| --- | --- | --- | --- |65| Signal | Dark console, live feed | The differentiator must be visible | Reads operational, not warm |66| Notebook | Warm paper, annotations | Trust through scholarship | Quietest, least product feel |6768Never mark a favorite in the table. Asked directly, answer from the product's69differentiator and audience, not which file was most fun to build.7071## Promote, then delete7273On a choice: port the winner into the real stack (components, tokens, routes),74keeping the contract's copy and the floor's checks. Delete the losing files and75the gallery unless the user asks to keep them — harness files that linger get76mistaken for product.7778## Outside references7980The project's own design system decides first. These fill the gaps it has not81decided, most often when picking direction worlds. Fetch the real content82before citing a reference — if a page cannot be read, say so and design from83the system's own tokens. Never imitate from memory. Report which reference was84used and what it changed.8586Style directions:8788| Reference | What it gives |89| --- | --- |90| styles.refero.design | 2,000+ named `DESIGN.md` style directions from real product sites — colors, type, spacing per world. The fastest way to name a direction. |91| VoltAgent/awesome-design-md | Open-source `DESIGN.md` collection for ~74 brands, MIT. |92| minimal.gallery | Curated minimal-site gallery; tag pages (`/tag/saas/`, `/tag/startup/`, `/tag/one-page/`) for restraint baselines. |9394Components and interaction (full list in `ui-micro-interactions`):9596| Reference | What it gives |97| --- | --- |98| 21st.dev | React + Tailwind component registry with an MCP; free usage limit — say what you are looking for before calling. |99| component.gallery | 60 components across ~95 design systems. Check how mature systems handle a component before inventing one. |100| microkit.co | MIT micro-interactions on the shadcn registry. |101| aicss.dev | Agent-UI component taxonomy (thinking, tool calls, streaming, approvals). Pro source — use the taxonomy, not the code. |102103Motion and video:104105| Reference | What it gives |106| --- | --- |107| kinetics.colorion.co | Spring-physics motion recipes; copy CSS, React or an AI prompt. |108| whatships.com | Launch-video library; tile a reference's frames to study pacing. |109| hyperframes.dev | Renders an authored HTML sequence to MP4 for demo videos. |110| impeccable.style | `polish`/`distill`/`bolder` review commands for "still feels off" work. |111| jakubkrehel/skills | `better-*` interface skill collection — `npx skills add jakubkrehel/skills`. |112113## Before you finish114115| Mistake | Fix |116| --- | --- |117| Variants differ only in accent or font | Merge them; write a different world |118| Builders saw each other's output | Rebuild the converged ones blind |119| Judged from thumbnails or file:// | Serve over HTTP, review at real size |120| A variant clips words at 390px | Fix the floor before it enters the gallery |121| Table marks a favorite | State each world's cost; the user picks |122| Harness survives promotion | Delete it in the promotion commit |