Back to all posts

What's Actually Inside the Claude Code Starter Pack I Used to Ship Liftkit

·7 min read

Most "Claude Code starter pack" downloads I've opened in 2026 are a folder of generic prompts you could've written yourself in five minutes. They sell vibes, not structure.

The starter pack I used to ship liftkitstudio.com isn't that. It's the version refined on a real product — the one I used to go from a Notes-app sketch to a live Next.js + LemonSqueezy commerce site over three evenings, and the same one I'm shipping to other solopreneurs now. This post walks through what's actually in it, file by file, in the order Claude Code reads them, and what shipping a real product taught me about which pieces matter most.

If you're evaluating whether a starter pack is worth $99 or whether you should just glue one together yourself, this is the post. By the end you'll know exactly what you'd be skipping by going DIY — and whether that skip is worth the time.

What it actually took

The starter pack is small on purpose. Here's the inventory of what's actually in the zip you download:

  • 1 CLAUDE.md (~3.6 KB). The project constitution Claude Code reads on every single turn.
  • 8 numbered phase skills (0-start, 1-explore, 2-define, 3-validate, 4-scope, 5-spec, 6-prototype, 7-build). One per phase of the framework, each shipped as a Claude Code skill with a fixed input/output contract.
  • 4 utility skills (decide, progress, feedback, improve). Cross-cutting skills you call from any phase — for triaging stuck decisions, checking progress, capturing feedback, and applying it.
  • 3 build templates (product-claude-template.md, project-progress-template.md, build-brief-template.md). The artifacts the build phase fills in.
  • 1 initiative output template + 1 README + 1 .gitignore.

That's it. About 17 markdown files. No bloat, no padding, no auto-generated boilerplate. By comparison, the most-marketed competitors ship 200-1000+ "agents" or "commands" — and you'd never realistically test all of them, let alone trust which ones survived a real production run. Less is more — and you can actually read every file in the pack in under an hour before you start.

What the pack does NOT include: a Next.js scaffold, a deploy stack, design tokens. Those decisions are yours — the pack is the AI-guided framework (the brain), not the final code (the body). You bring your stack; the skills walk Claude through using it.

The pack ships as a single download (no telemetry, no auth wall, no email gate after purchase). You unzip, you read, you start.

What's inside the template

Here's the order Claude Code actually reads these skills when you start a project, and why each one is positioned where it is.

Step 1 — drop in CLAUDE.md. This is the file Claude reads on every single turn. The starter pack ships a templated version you fill in with your stack, brand, and guard rails before any code is written — the boring part most "AI builder" tutorials skip. The CLAUDE.md is where roughly 80% of the leverage of the entire pack lives. Get this wrong and every subsequent skill produces mediocre output. Get it right and Claude stops asking basic questions and starts shipping.

Step 2 — invoke 0-start. The entry skill. It walks Claude through reading your CLAUDE.md, confirming what kind of project you're building, and telling you which phase to enter first (almost always phase 1, but 0-start exists for the cases where you're resuming a paused project mid-flight).

Step 3 — work through 1-explore → 7-build. Each numbered phase is a Claude Code skill that takes its predecessor's artifact as input and produces its own artifact as output. 1-explore asks YOU four questions before any building starts: who is the buyer, what are they searching for, what's the simplest thing that would solve their problem, and what's the kill-gate (the one signal that means don't build this). 2-define writes the one-pager. 3-validate is the painful one that kills bad ideas. 4-scope cuts to the smallest shippable thing. 5-spec writes a 5-minute-readable technical spec. 6-prototype gets you to a clickable, branded thing. 7-build is the boring deploy work — your stack, your DNS, your payments, your analytics. Each phase's output is concrete; each is reviewable in isolation.

Step 4 — call utilities when stuck. The four utility skills (decide, progress, feedback, improve) are designed to interrupt the linear flow. Stuck deciding between two scope options? decide triages. Lost track of where you are? progress shows the picture. Something feels off? feedback captures it; improve applies the captured items as a batch. These are the skills that prevent solo founders from spinning out at hour 14 of a build.

Step 5 — diverge. By the end of phase 7, you should be off the rails. The starter pack is scaffolding, not a cage. The CLAUDE.md you customized in step 1 stays — it's still your constitution. Everything else is meant to be ignored once you don't need it anymore.

What broke (and how I fixed it)

Three hours before launch, I killed my own geo-pricing.

Liftkit was originally built to detect the visitor's country and show prices in their local currency — so a buyer in Bogotá saw "400.000 COP" and a buyer in Mexico City saw "1.890 MXN." I'd spent real time wiring up geo.ts to read Vercel's request headers, mapping countries to currencies in pricing.ts, and getting the LemonSqueezy checkout to honor the local denomination. It worked. It looked smart.

Then, three hours before I planned to push live, I sat down and looked at what a Colombian buyer would actually see: 400,000 pesos. Six figures. Even though it's roughly $99 USD, the visual of a six-digit price tag is psychologically heavier than a two-digit dollar sign — especially for a digital product the buyer hasn't seen anyone else recommend. Conversion was going to lose to sticker shock that wasn't even real.

So I tore it out. Thirty minutes of work to revert: kill the country-detection branch in pricing display, set everything to USD, leave geo.ts in the codebase for a future use case (it's still there, dormant, in case I need it for tax or shipping logic later). LemonSqueezy already handles tax-inclusive checkout per region — that part stays automatic. What changed was the display: every visitor now sees $99 regardless of where they're from, and the conversion psychology aligns with the price reality.

The lesson isn't "geo-pricing is bad." It's that the version of the framework I'd written before this moment didn't have a "test the price visually in the buyer's eyes, not just the API" check. It does now. The starter pack improves every time something breaks (or almost-breaks) in production. The bloated competitors have more prompts because none of theirs have been pressure-tested by an actual launch.

The takeaway

If you're a solopreneur evaluating whether to spend $99 on a starter pack or DIY one yourself, here's the honest math.

Building this from scratch took me roughly two weeks of evenings — checking what the existing "AI builder" content actually delivered, drafting a phase framework that didn't read like generic productivity advice, testing each skill against a real project, rewriting the ones that didn't survive contact with production. That's the cost when you have my context. If you're starting cold and don't already know which skills survive a real ship, the realistic estimate is at least one solid week of focused work — and that's assuming you don't get distracted by the 200-skill bloated alternatives and end up rebuilding theirs instead of yours.

Buying the pack costs $99 and roughly four hours of customization to make it yours. The arbitrage isn't subtle.

But the actual reason to buy isn't the hours saved. It's that the version you'd write yourself, on your first project, won't be the version that survives contact with production. The pack's value is the iterations baked in: the geo-pricing revert from above, the kill-gate questions in phase 1 that catch projects that shouldn't ship, the explicit hand-off contracts between phases. You're not paying for skills. You're paying for the mistakes already made — so you don't have to make them too.

Get the Claude Code Starter Pack — $99

Built once, refined across multiple ships, sold from the site it built.

Keep reading