Decisions Claude can’t un-commit you from
You can vibe-code your way to a working prototype. You cannot vibe-code your way out of a bad infrastructure decision made in week one.
Most of this course is about how to work with Claude. This chapter is about the small set of choices that sit upstream of that work, where Claude will happily help you ship faster than you can untangle the trade-offs. None of these decisions can be reversed by a regeneration from the spec. They need a minute of human thought before you commit.
This is the short list of questions you want answered before you pick.
Platforms: where does this live?
Decide first: web app, native app, chat interface (a Claude Skill or GPT action), or some mix.
- Web is the most portable and the easiest to iterate on.
- Native gets you OS integrations, notifications, and distribution through stores, at the cost of release friction.
- Chat interface is the newest option; low friction for users but a limited UX surface.
Many products end up multi-platform. Starting on one and adding the others later is almost always easier than starting on all three.
Auth: who does it and how much can you own?
- Google / Apple / social auth: fastest setup, users already have accounts, no password reset pain. Lock-in is soft.
- Your own auth (email + password, magic link, passkey): full control, more surface area to maintain, more security responsibility.
- Third-party identity providers: Auth0, Clerk, WorkOS, Stytch. You get enterprise features (SSO, SAML, RBAC) without building them. You also get vendor pricing that escalates and occasional lock-in on session formats.
If you’ll serve enterprises eventually, pick an auth provider that can do SAML/SSO from day one; retrofitting it later is painful.
Database hosting: the decision you’re stuck with longest
- Neon, Supabase, PlanetScale: managed Postgres or MySQL variants with good developer UX. Quick to start, cost-effective at small scale.
- RDS / Cloud SQL: classic managed options from AWS / GCP. Flexible, enterprise-friendly, more config.
- Self-hosted Postgres: the old reliable. Max control, max responsibility.
Migration strategies are almost always hand-rolled. Whatever provider you pick, assume you’ll eventually move off it. Keep schemas portable, avoid vendor-proprietary features where you can, and factor migrations out of your app code.
App-layer hosting
- Vercel: the default for Next.js; opinionated, fast to deploy, limits get expensive at scale.
- Cloudflare Pages / Workers: great for static and edge-served apps; growing capability for full-stack.
- Fly.io / Render / Railway: generalist platform-as-a-service options for apps that outgrow Vercel’s constraints.
- Your own AWS / GCP: max flexibility, max ops burden.
Watch for pricing cliffs: many hosts are cheap until a specific usage threshold, then steep. Understand the pricing curve before committing architecture to a host’s primitives.
I/O services: email, SMS, notifications
- Email: Resend, Postmark, SendGrid, SES. Pick one that handles deliverability well; pricing is usually second-order.
- SMS: Twilio, MessageBird. Gritty with number provisioning and carrier rules; plan for that complexity.
- Push notifications: Apple / Google native stacks; Expo if you’re cross-platform.
The trap here is usually late integration: shipping a product, then adding transactional email at the last minute, then discovering your host doesn’t play well with your email provider.
Observability: see what’s happening
- Error tracking: Sentry is the default; good free tier, integrates everywhere.
- Logs and metrics: New Relic, Datadog, Grafana Cloud. Pricing varies wildly; tail-sampled logs are a common gotcha.
- Application performance: often built into the metrics tool; sometimes a separate vendor.
Start simple: Sentry + console logs is fine for most apps under 1,000 users. Upgrade when you actually need the extra signal.
User analytics
- PostHog: self-hostable, generous free tier, good product-analytics feature set.
- Amplitude / Mixpanel: deeper analytics, pricier at scale.
- Google Analytics: marketing-focused, lighter on product-analytics signal.
The trap: events get over-tracked and under-cleaned. Decide on a tracking plan before wiring anything up, or you’ll drown in noise within a month.
Ads and marketing platforms
Each platform (Meta, Google, TikTok, LinkedIn) has distinct APIs, attribution models, and reporting quirks. Picking one to start is fine. Picking three simultaneously without infrastructure to compare them is a trap.
A pragmatic default stack for small teams
For solo / small-team builders without specific requirements:
- Hosting: Cloudflare Pages or Vercel
- Database: Neon (Postgres) or Supabase
- Auth: Clerk or your platform’s native auth
- Email: Resend
- Analytics: PostHog
- Error tracking: Sentry
All of these have generous free tiers, work well together, and can scale you to small-revenue before you need to revisit.