Public decision log

Used Next.js 16.2.4 for the new build, with Turbopack + App Router.

We needed 4-locale SSG with sub-millisecond TTFB, content-driven (not CMS-driven), and Docker Swarm deployable. Next 16's `params: Promise<{...}>` async API is the right tradeoff: it forces correct async handling and unlocks streaming SSR later. Next 15 has the same shape but still in development. Going back to 14 would have meant missing 2 years of perf improvements.

Astro is great for content sites but our dynamic pricing + case-study pages would have needed client islands anyway. SvelteKit is faster but the team has 5 years of React muscle memory. Remix is React 18 only. Vite + RRR is 3x more code to maintain. Pages Router is dead.

All 42+ client sites run on a single VPS with Docker Swarm, Traefik v3.5, and CF orange-cloud proxy in front.

We have 1 engineer (me) running 42 client sites. K8s would mean a cluster, control plane upgrades, RBAC, ingress controllers, cert-manager, etcd backups — that's 3-5 days/month of yak-shaving per month. Swarm has the same declarative model with one binary, one `docker stack deploy`, zero cluster management. The downside is no auto-scaling, but we don't need it — the entire fleet peaks at ~30% CPU.

K3s still needs cluster maintenance. Nomad has a smaller community. docker-compose can't do zero-downtime deploys. Vercel/Netlify would mean 42 separate bills and 42 separate build pipelines. CF Pages requires CF Workers build system, less flexible.

All 28 services with reference prices, market low/high, and internal proof — on a public Google Sheet + on the company site.

When I started AI Whisperers, the #1 buyer objection was 'how much?'. The #1 buyer complaint after hiring us was 'I had no idea it would cost that'. Both come from the same root: hidden pricing. Publishing rates up-front filters out clients who can't afford us (good — wrong fit) and earns trust from clients who can (good — long-term ROI). The risk is competitors undercutting us. The counter-risk is looking like every other opaque agency in Paraguay.

Quote-only means a 30-min call for every prospect, even the ones who can't afford us. Tier-only is the worst — it's hard to compare and lets us hide the actual rates. Hourly-only is technically transparent but psychologically worse: a $5K quote looks like 50 hours × $100, which is intimidating.

All site copy is in `content/{en,es,nl,pt}/site.json` + per-section JSON files. No CMS. Git is the source of truth.

For a 4-locale, 44-page site with 28 service items, the CMS overhead doesn't pay back. Git gives us: (1) versioned content, (2) PR review on copy changes, (3) Claude/Cursor can edit JSON in seconds, (4) no vendor lock-in, (5) no separate deploy pipeline. The downside is non-technical users can't edit — but we have 2 people, both technical, and the client doesn't need to edit copy (they sign off on it before deployment).

Sanity/Contentful cost $200+/mo and add a separate API to debug. Strapi is just more YAML config to maintain. Decap CMS is good but adds a runtime. Markdown would work but JSON is more structured for our use case (pricing tables, testimonials, etc.).

Built a fresh Next.js 16 + Turbopack app, deployed to our own VPS, killed the Vercel deployment. Same DNS, different backend.

The old Vercel build was the 'Machines for the math. Humans for the magic.' tagline — circa 2025. By 2026 we'd moved past the AI-training pitch into the AI-engineering pitch. The old tagline was hurting us in sales calls. Iterating on the Vercel build would have meant: editing React components, hitting Vercel's API to redeploy, dealing with the Vercel-only edge functions, etc. Starting fresh on our own VPS gave us: full control, no per-function billing, security headers, Traefik middlewares, content-driven rebuild. Took 4 days end-to-end.

Vercel iteration kept the wrong architecture. Astro would have meant rewriting every interactive component. Hugo would have meant no SSG with per-locale content. Fresh Next on VPS won on all axes.