Local services · Bilingual digital identity for a two-person veterinary clinic
LOVET VETERINARY CLINIC IDENTITY
A two-person veterinary clinic competed for local attention against bigger practices with no ad budget and no digital identity. The build made "we're only two people" the pitch, not the limitation, and backed it with a 99/100 PageSpeed build.
- Role
- Solo: brand direction, design system, component and i18n architecture, copy, technical SEO and performance, with Claude Code as the build tool
- Timeline
- 2026 · live on lovetclinicaveterinaria.es
- Client
- Lovet · Veterinary clinic in Sant Feliu de Llobregat. Carme, veterinarian, and Xavi, veterinary assistant
- Stack
- Next.js static export, Tailwind CSS, sharp-based image scripts, Web3Forms, Vercel hosting
- 99 PageSpeed performance on mobile, 100 on desktop
- 0 Bytes of already-live Catalan content changed
- 3 Independent cohesion audits before any component got consolidated
A two-person clinic competing against bigger names
Lovet is a real veterinary clinic in Sant Feliu de Llobregat, run by two people: a veterinarian and a veterinary assistant. With no website carrying its own identity, it competed for local attention against bigger clinics with a stronger digital presence, no ad budget, and no acquisition channel beyond word of mouth and local search.
The business problem was clear: a small clinic needs its website to do the trust-building work that a large clinic gets for free from headcount or years in business. The product problem was different. This wasn't about building "a veterinary clinic website" in the generic sense. It was about turning what could read as a limitation, just two people, into the pitch itself.
What made it hard
- Starting point was a template. The project began on Obliq, a generic Next.js/Tailwind template. The real risk was shipping a surface-level rebrand that still read as a template with a different logo
- Static hosting, no backend. output: 'export' with unoptimized images, chosen for simple hosting proportional to the business, rules out next/image from the start. Every image optimization has to be solved by hand
- Limited source material. The founders' real photos arrived at 480px. Any layout that enlarged them pixelated
- Bilingual added after the fact. Catalan was already live and stable when the decision came to add Spanish. Touching the architecture without risking a single character of already-indexed Catalan content was a self-imposed constraint, not a technical requirement
- Zero budget for deployment mistakes. The Vercel project had framework: null, so a mis-executed deploy served the public/ folder as the web root. It broke production once
What I owned, and what the client decided
As the sole designer and developer on the project, with Claude Code as the working tool, I owned it end to end: brand direction and design system, component architecture, i18n architecture, copy, technical SEO, performance and the deployment pipeline itself.
- The client decided: the underlying positioning ("it's just the two of us, and that's the advantage"), which legal content to use, dropping the "What's New" section and the visual skip-link, accepting the Spanish legal text as an unreviewed translation, and the brand naming ("LOVET" in capitals)
Less brochure veterinary clinic, more "it's just the two of us"
The first draft of the copy read like a generic clinic: "comprehensive care", "qualified professionals". The real pivot was treating the small scale as the differentiator: "At Lovet there are only two of us: a veterinarian and an assistant who know you by name." That line, not a services list, is the first thing any visitor reads.
Bilingual without touching a single byte of what was already live
Instead of merging Catalan and Spanish under one shared [locale] route, the architecture uses a transparent (ca) route group, Catalan stays at the root exactly as it was, and a separate [locale] tree just for es. To verify it, I built verify-ca.mjs: it normalizes each build's non-deterministic noise (webpack hashes, React hydration comments) and diffs the resulting Catalan HTML byte-for-byte against a baseline. Any real content change gets caught; a harmless rebuild doesn't. That turned "did I break the Catalan site?" into npm run build && node scripts/verify-ca.mjs, not a manual review.
Root-causing a real CLS, not masking it
A PageSpeed audit flagged a ~0.143 CLS on desktop. The cause wasn't cosmetic: the body font (Elms Sans) loaded with font-display: swap and no fallback font with matching metrics, so every line of text reflowed the instant the web font arrived. The fix was generating a local fallback font, Arial plus size-adjust, ascent-override and descent-override calculated for that specific file, not disabling the swap or papering over the jump with opacity tricks.
Consolidating components only once the duplication was real
Three cohesion and visual-quality audits, 40, 34 and 36 findings across separate rounds, found that services, facilities, plans and specialists used 4 to 6 near-identical variants of the same card, CTA band or icon set. They were consolidated into single primitives (person-card, cta-band, icons, trust-chips) only after confirming 3 or more real copies existed, not as speculative abstraction up front.
Three smaller calls that add up
- Manual image pipeline, not next/image. With simple hosting as a business requirement, the alternative was targeted sharp scripts to crop and generate responsive variants only where the real weight justified it: the home carousel served 1024px images inside 288 to 384px boxes, so 640px derivatives were generated. Trade-off accepted: every new image goes through a script before upload, not an automatic API
- WhatsApp and a light form, not a booking backend. The clinic runs on phone and WhatsApp in practice. Building an appointments backend would have solved a problem they don't have. The contact form uses Web3Forms over fetch, with a honeypot against spam and a phone/WhatsApp fallback if the API key is missing
- No invented legal data. While the tax ID, professional college number and accountant's details were still pending, the legal content showed a visible amber marker ("Fill") instead of a "reasonable" placeholder. It was replaced with the real data as soon as the accountant delivered it
When the source material dictates the layout, not the other way around
The clearest case of real iteration, not cosmetic polish, is the founders' profile layout.
- V1, large card. With source photos at only 480px, scaling them into a large card format pixelated them
- V2, alternating editorial. An asymmetric column grid ([340px_1fr]) with order-last to alternate the photo side between the two founders. Real bug: order reorders the element visually in the DOM but doesn't swap the column width, so the photo kept landing in the wide track and stretching the same way
- V3, twin columns. Final layout: grid-cols-2 with no alternation and no order, portrait capped at a max width (roughly 340 to 440px) that doesn't force upscaling. The two founders also end up visually equal, which was the client's actual ask
A client who made real product calls
The client wasn't a passive stakeholder. They made explicit product decisions that changed scope: dropping the "What's New" section before it was ever built, removing the accessibility skip-link because it visually overlapped the logo (the underlying id="main" anchors stay in place in case it's revisited), and knowingly publishing the Spanish content as a faithful translation of the Catalan without a lawyer's review in that language.
99 on mobile, 100 on desktop, measured 12 September 2026
Measured with PageSpeed Insights on www.lovetclinicaveterinaria.es: 99 performance, 100 accessibility, 100 best practices, 100 SEO on mobile, and 100, 96, 100, 100 on desktop.
Real Core Web Vitals: LCP 2.2s mobile, 0.5s desktop. CLS 0 mobile, 0.033 desktop. TBT 0ms mobile, 20ms desktop.
What's verified, and what I won't claim
The 96 accessibility score on desktop isn't a number polished for the occasion: PageSpeed flags insufficient contrast on the "ES" link in the language switcher, text-deep/60 over the header background. It's a documented open item, not a hidden one.
What I can't claim: there's no baseline from the previous WordPress site on analytics, speed or conversion, so there's no "improvement over before" figure. There's also no Search Console data, no GA, and no interviews with real users in the material available. The documented impact is technical, performance, accessibility, on-page SEO, and product, features shipped, not business.
The lesson wasn't "audit more"
The project went through three independent audit rounds before any area was called done, and each one found real duplication the previous rounds missed, because they looked page by page, not at the whole system. The lesson isn't to audit more. It's that component consolidation only makes sense when you look at the whole, not the isolated piece.