Back to all posts

The Claude Code Template I Used to Ship a Real Site as a Solopreneur (in 3 Hours)

·7 min read

The first deploy worked. That was the part I didn't expect.

I'm a solopreneur with a 9-5. I spent years at Rappi (one of Latin America's two unicorns at the time), then moved to Addi (a Colombian fintech, close to unicorn now). Liftkit was supposed to be a weekend toy — a marketing site for a digital product I'd been thinking about for months. Two hours of planning, three days of build work spread across evenings, and on April 16, 2026 the site went live on Vercel.

Eleven days after launch, 50+ organic users from outside my home country have found their way to liftkitstudio.com through Google — no ads, no audience, no email list, no Twitter following, no local network advantage. The site you're reading this on was built by the Claude Code template I'm about to describe. That's the only credential that matters.

What it actually took

Here's the actual decomposition of the build, reconstructed from commit timestamps in the Liftkit repo:

  • Planning (~2 hours, day zero): No figma, no prototype, no design system handed down. Just a few online references for sites I liked the feel of, a CLAUDE.md I wrote on a napkin (literally — Notes app), a sketched file structure, and a stack decision. No code yet — just decisions. This is the part most "build with AI" videos skip and it's also the part that determines whether the next phase takes 3 days or 3 weeks.
  • Build day 1 (~1 cumulative hour, evening): Claude Code read the CLAUDE.md and scaffolded Next.js 16 + App Router + Tailwind + shadcn/ui. Components, layout, design tokens applied from the prototype phase.
  • Build day 2 (~1 cumulative hour, evening): MDX content directory, two pricing tiers, LemonSqueezy overlay checkout wired up with tax-inclusive pricing for three currencies (USD, COP, MXN — buyer profile is global solopreneurs, not just US).
  • Build day 3 (~1 cumulative hour, deploy): Vercel + Porkbun DNS + GA4 + Search Console. The only thing that broke during deploy is the thing I'll tell you about in the failure section. Spoiler: it was DNS. It's always DNS.

Two hours of planning, three days of evenings, ~3 cumulative hours of active build work. One person, no team, no investors. Total cost to ship: ~$15 (domain) plus the Claude Code subscription. Live since April 16, 2026. As of writing this (April 27), 50+ organic users from Google — measured outside my home country, so it's not local-network noise.

What's inside the template

Here's what's actually inside the Claude Code template that produced this site, in the order Claude reads it:

1. CLAUDE.md (the constitution). This is the file Claude reads on every turn — your project's rules, stack, voice, and guard rails. Liftkit's CLAUDE.md sets the brand, the design tokens (copied from the prototype phase), the file structure (App Router, MDX in src/content/, no database, no auth), and a list of things Claude is never allowed to do (like adding a state-management library or building user authentication). The CLAUDE.md is where 80% of the leverage lives. A good CLAUDE.md means Claude doesn't have to ask basic questions — it just builds.

2. The 7-phase framework (the path). Most "build with AI" content treats AI like a code generator. The template treats it like a co-founder with a roadmap. The phases are: Explore (decide what to build), Define (write a problem brief), Validate (kill bad ideas before scoping), Scope (cut to MVP), Spec (technical spec a human can read in 5 min), Prototype (clickable, branded), Ship (the boring deploy work). Each phase has a concrete artifact at the end — not a vibe, an output.

3. The prompt library. A folder of .md files, each one a battle-tested prompt for a specific job: "write the landing page hero," "scaffold the pricing config," "wire up GA4 events with the right scroll-depth thresholds." When you're tired (and at hour 2 of a build, you will be), you don't think — you copy the prompt, fill the variables, paste, ship.

4. The kill-gate checklist. This is the unglamorous one. Before the framework lets you advance from one phase to the next, it forces you to answer three questions: Is there a real buyer? Can they find me? Will they pay? If any answer is "I don't know," you don't get to advance. Most solo projects die because someone scoped before validating. The template prevents that by structurally refusing to.

5. The starter repo (the artifact). A pre-configured Next.js 16 + Tailwind + shadcn + LemonSqueezy + GA4 setup with placeholders where your content goes and zero placeholders where the boring infra would be. You clone it, point it at your domain, drop your design tokens, and start writing content.

That's it. Five components. Nothing exotic.

The thing most "AI builder" content gets wrong: they treat the AI as the leverage. The AI isn't the leverage — the AI is a commodity. Anyone can prompt Claude. The leverage is the opinionated structure around the AI: a CLAUDE.md that encodes how YOU make decisions, a phase framework that forces validation before scoping, a prompt library where each prompt is the version that actually worked instead of the version that read well. That's the asset. That's what's hard to copy. The template is just the format that ships those opinions to you.

What broke (and how I fixed it)

Tonight (Sunday, ~9 PM), I clicked the wrong button in Porkbun and took liftkitstudio.com offline for ~10 minutes.

The button said "Fix DNS." I assumed it would fix DNS. It rewrote my A records to point at Porkbun's parking page. The site I built in 3 hours started serving "A Brand New Domain!" to anyone who landed on it. I noticed because I was literally testing the Search Console wire-up for a different post when the page 404'd.

Recovery took 8 minutes once I knew where to look: edit the ALIAS and wildcard CNAME at Porkbun, point them at cname.vercel-dns.com, wait for propagation, hard refresh. Vercel's edge picked up the binding within 90 seconds. No customers lost (it's a Sunday night and the site is small).

The reason the recovery was 8 minutes and not 8 hours is the same reason I'm able to maintain this site at all: there are five components in the stack. Domain (Porkbun), DNS (Cloudflare-backed via Porkbun), edge (Vercel), payments (LemonSqueezy), analytics (GA4 + GSC). That's it. When something breaks, there are five places it could be. You can hold five things in your head while you debug at 9 PM after a long day. You cannot hold fifty.

Real lesson: in a single-operator product, the same person ships AND fights the fires. The template won't save you from clicking the wrong button. What it does is keep your stack so simple that when you DO click the wrong button, you can fix it in 8 minutes instead of 8 hours. Five components, not fifty. That's the whole bet.

The takeaway

The Claude Code template doesn't make you faster at writing code. Claude already does that.

What the template makes faster is the boring decisions: which stack, which payment provider, which folder layout, which prompts to fire in which order, which gates to enforce before you advance. Decisions made once, batched into a CLAUDE.md, frozen. You stop spending willpower on infra and start spending it on the thing only you can do — the brand, the words, the offer.

If you're a solopreneur with a 9-5 and 6 hours of weekend energy, the template isn't a productivity hack. It's the thing that decides whether your next idea ships or sits in a Notion doc for another six months.

Get the Claude Code Template — $99

Built once, shipped in three evenings, drawing organic search traffic from week one. Receipts in the URL bar.

Keep reading