Are we stuck with the tools of 2024 because that's what the models were trained on? Or as they improve, do choices in fact widen? AI assistance quietly changed the calculus of stack choice: the best-documented, most-discussed technologies got a productivity multiplier that obscure-but-elegant ones didn't. Alongside it, not-invented-here syndrome has also multiplied: "Why pay for a SaaS when I can build it myself for nothing?"
I can share the stack we've settled on at Poteris for scalable apps on nonprofit budgets, and why each piece earns its place: Cloudflare for hosting/edge/DNS, Neon for serverless Postgres, Drizzle as the ORM, Better Auth for authentication, and Next.js or Svelte as the framework (a debate that's still live in our team).
And of course, "I can build it myself" continues to be a fallacy and a time-sink in all but the simplest cases. Any tool with a little meat to it benefits hugely from the fixes for all the bugs and pain points people have found over a period of use, and continued maintenance and improvement. Off-the-shelf remains the default choice: focus on the bits you're interested in, and piggyback off others for the rest (although personal/one-off tools with a very specific use can be a different matter).
The framework for making these choices: minimise the number of things that can wake you up at night, maximise what AI tooling can help with, let others do the heavy lifting where you can, and keep your escape routes open.
Key question/s: What should you build, what should you buy, and what should you let the models' training data choose for you?
Reading/links
- Choose Boring Technology – Dan McKinley's essay on stack choice
- In Defense of Not-Invented-Here Syndrome – Joel Spolsky's classic steelman for building it yourself
- Next.js · Neon · Drizzle ORM · Better Auth · Cloudflare – the Poteris default stack, all with generous free tiers for prototypes
- Svelte – the other side of our still-unsettled framework debate