# GGQ UI demos v2: Vercel Skills A/B — narration (first-person)

Generated (UTC): 2026-01-22T09:32:53Z

I wanted a clean, apples-to-apples test of whether Vercel’s “Skills” could meaningfully improve the GGQ UI demos layout and UX without me going off-script and doing a full redesign.

So I treated it like a production A/B:
- A = baseline (already published, do not touch)
- B = skills-driven fork (new publish, isolated)

The goal wasn’t “make it prettier by taste” — it was: can a codified checklist catch real problems and help me ship a measurably better front-end quickly?

## Act 0 — I locked the baseline so I couldn’t cheat

First I confirmed the baseline pack is reachable and committed to the rule: no edits to it.

Baseline (A):
https://infrafabric.io/static/hosted/review/ggq/ui-demos/2026-01-22-v2/index.html

Skills fork (B):
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/index.html

That constraint was surprisingly freeing: once “don’t touch A” is non-negotiable, every change naturally becomes a controlled experiment.

## Act 1 — I brought in Skills, but I refused to let it pollute the machine

I expected the first friction to be npm/npx sprawl (global installs, random caches, “where did this file come from?”).
So I ran Skills with an isolated HOME and cache.

This part was genuinely nice: the “skills” concept is basically “install a procedure”, and the procedure itself is just a markdown contract.

Delight: it felt like adding a new checklist to the kit without installing a whole framework.

Friction: the moment you run anything from the internet, reproducibility anxiety kicks in — so I pinned the source.

## Act 2 — I pinned the exact guideline bytes (so the audit is repeatable)

Instead of trusting “whatever GitHub serves today”, I pulled the guideline source used by the skill and recorded its sha256.

Pinned guideline source:
https://raw.githubusercontent.com/vercel-labs/web-interface-guidelines/main/command.md

Pin record:
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/report/guidelines_pin.txt

Skill sources captured:
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/skill_source/SKILL.md
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/skill_source/web-interface-guidelines.command.md

Delight: once pinned, it stops being “Vercel says” and becomes “these exact bytes said”.

## Act 3 — Audit baseline (A): the skill did what I wanted (it was unromantic)

I ran the guidelines against the static HTML/CSS/JS in the GGQ demo pack.

Baseline findings report (95 items):
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/report/skills_audit_findings.txt

The pattern was consistent (and very fixable):
- lots of CLS risk: images missing explicit width/height
- a11y gaps: search inputs with only a placeholder (no label/aria-label)
- polish items: focus states and motion preferences

Delight: the output was crisp — file + line + issue — no sermons.

Frustration: it was also repetitive (the same two issues repeated across demos), which is exactly the kind of work humans hate doing by hand.

## Act 4 — Visual check: I got reminded that screenshots can lie

I generated baseline screenshots so we have a “what the user sees” reference.

Desktop A/B:
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/screenshots/AB_home_desktop.png

Mobile A/B:
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/screenshots/AB_home_mobile.png

Here’s a subtle but important friction:
- headless Chromium can capture a half-rendered state if scripts haven’t finished
- if networking is disabled (or slow), remote images don’t load and the UI can look “broken” even when it wouldn’t be in a real browser

So I increased the virtual time budget and allowed the page to settle before capturing.

Delight: once I treated screenshots like evidence to control, they became useful.

Frustration: Chromium spammed unrelated warnings (PHONE_REGISTRATION_ERROR). Annoying, but harmless.

## Act 5 — Fork (B): I applied fixes like a mechanic, not a designer

I made a clean fork of the baseline demo files and applied only “skills-driven” changes — the smallest fixes that remove the findings.

What I changed (high level):
- added `aria-label` and a stable `name` to the search inputs (so it’s not placeholder-only UX)
- added explicit `width`/`height` on images (reduces layout shift)
- added strong `:focus-visible` styling (keyboard users aren’t punished)
- honored `prefers-reduced-motion` (don’t force motion when the user asked for calm)
- kept everything else intentionally boring (no big layout rewrites)

Then I reran the audit on the fork.

After report (0 findings):
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/report/skills_audit_findings_after.txt

Delight: going from 95 → 0 is a satisfying “green checkmark moment”.

Frustration: while automating the edits, I hit classic scripting papercuts:
- I accidentally wrote literal strings like `$(date -u ...)` into a file because I quoted a heredoc the wrong way.
- I tripped a `sed` regex error (“Unmatched (”) while trying to do a fast replacement.

It’s the same feeling as dropping a screw in grass: the fix is simple, but the interruption breaks flow.

## Act 6 — Packaging + publish: I kept the experiment shareable

I bundled the fork as a tarball so you can diff/inspect it locally.

Pack tarball:
https://infrafabric.io/static/hosted/review/ggq/ui-demos-skills/2026-01-22-v2-skills/ggq-ui-demos-2026-01-22-v2-skills-pack.tar.gz

I left a trail of what was used and what changed so another session can reproduce the result rather than “trust my vibe”.

## Epilogue — What I learned (the real A/B result)

Skills helped most with the kind of work that is:
- repetitive
- easy to miss
- easy to verify after the fact

In this run, the biggest practical wins were accessibility basics and layout stability (CLS).
It didn’t magically invent a new design language — and that’s good. It behaved like a disciplined checklist.

If we want wow-level visual improvements, Skills is a strong floor, not the ceiling: we can layer an explicit GGQ design system / component pass on top, but keep the skills audit as a non-negotiable gate.
