Why this default pays off

Server Components ship zero JavaScript to the browser for the parts of a page that don't need it. On a marketing page, that's often most of the page — headings, copy, and images render as plain HTML.

A concrete example

A pricing table's copy and layout can stay server-rendered; only the currency toggle, if one exists, needs to be a small client island around it.

// Server Component — no "use client" needed
export function PricingCard({ tier }: { tier: Tier }) {
  return (
    <div>
      <h3>{tier.name}</h3>
      <p>{tier.price}</p>
    </div>
  );
}

Where we still reach for client components

Anything with autoplay, drag interactions, form state, or scroll-based animation. The goal isn't "zero client components" — it's not paying the client-JS cost for content that never needed it.

Start with the workflow, not the model

Before choosing an API or a prompt strategy, write down the exact decision a person currently makes by hand. That's the task worth automating first.

Keep a human in the loop where it matters

For anything customer-facing or irreversible, route low-confidence outputs to a person instead of guessing. This is usually a small amount of extra engineering that saves a lot of trust.

What "low confidence" looks like in practice

Often it's as simple as a threshold on a classification score, or asking the model to flag its own uncertainty as part of its output schema.

Signs an MVP is the right call

You're not yet sure real users want this, pricing is unvalidated, or the core workflow itself might change based on early feedback.

Signs a fuller build is worth it up front

You already have paying customers waiting, or the integration/compliance requirements are substantial enough that a thin version wouldn't be usable at all.

Design around one question at a time

Ask what decision this screen needs to support, then cut every metric that doesn't inform that decision.

Default to summaries, let detail be optional

A drill-down interaction is cheap to add and keeps the default view calm.

Pick metrics tied to what users feel

Largest Contentful Paint and Interaction to Next Paint track much closer to "does this feel fast" than a raw Lighthouse score does.

Automate the check, don't rely on memory

A budget that isn't enforced in CI quietly erodes the first time someone's in a hurry.