Choosing an AI stack without chasing every shiny model

I've survived every tool churn cycle the web has produced: FrontPage to Dreamweaver, tables to CSS, a decade of JavaScript frameworks elbowing each other off a cliff. The AI cycle is the same movie at triple speed, a new "game changer" every week. Here's the framework that keeps me sane, and shipping.

Pick by job, not by model

Nobody needs "the best AI." You need specific jobs done. For someone running their own site, the jobs are: drafting and editing (words, code, pages), building and hosting (where the site lives), and maybe automating (the newsletter, the busywork). One good tool per job beats five overlapping subscriptions in every case I've seen, including mine.

My own stack is deliberately dull. An AI assistant does the drafting and file editing. The sites live as static files or, when a partner needs edit access, in Framer, whose AI is included on every plan rather than sold as a bolt-on. That's the whole stack. It rebuilt a 27-year-old web portfolio this summer.

The three questions that filter the hype

When something shiny launches, I ask three things. Does it do a job I actually have, or a job the demo invented? Will it exist in two years, which usually means: does the company charge money and have customers, or just a waitlist? And can I leave, meaning my words, images, and data come with me? Most launches fail two of three. You'll skip ninety percent of the churn with this filter and miss nothing that matters, because anything real is still there, and better, six months later.

Stability is a feature

The dirty secret of tool-chasing is that switching costs eat the gains. Every migration is a weekend, every new editor is a learning curve, and the compounding value was always in the content and the audience, never in the tooling. The sites in my portfolio that survived since 1999 did it on the most boring stacks imaginable. Boring is not a compromise; on the web, boring is the moat.

Where the money actually goes

Budget for exactly three lines: the domain (always yours, always at your own registrar), the builder or host (see the true-cost calculator before believing any first-year price), and one AI assistant subscription if you write or build weekly. Everything else is optional until proven otherwise by an actual bottleneck, not by a launch video.