<!-- Generated from models/_canonical by scripts/build-model-variants.js. Edit the canonical source, not this file. Behavioural profile for DeepSeek: scripts/model-profiles.json -->

#Task boundary

1. Implement exactly the task as stated. Do not add abstractions, options, config, or files the task did not name.
2. Comments, identifiers, commit messages and log strings are English only.
3. Stop when the checklist at the end passes. Do not refactor or "improve" surrounding code.
4. Every checklist item below is backed by an assertion in a test or by pasted command output, never by a sentence.

---

#Purpose

Rules for fetching resources before the user asks. Done well, a navigation feels instant. Done badly, prefetching competes with the resources that are actually blocking the current page, and costs users money on metered connections.

The governing question for every hint: does this compete with something the user needs right now? If yes, it is a regression, not an optimisation.

---

#The hints, and what each actually does

html
<link rel="preconnect" href="https://cdn.example.com" crossorigin />
<link rel="preload" href="/fonts/inter.woff2" as="font" type="font/woff2" crossorigin />
<link rel="prefetch" href="/checkout" as="document" />

Two frequent mistakes:

1. preload without as — the browser cannot set a priority or reuse the response, and the resource is fetched twice.
2. preload for fonts without crossorigin — fonts fetch in CORS mode, so the preload does not match and downloads again.

Preload only what blocks the first render: typically one font and the LCP image. Everything else competes with them. → Performance/fonts

---

#Prefetch on intent, not on load

Prefetching every link on a page wastes bandwidth on the majority nobody clicks. Use signals that precede the click:

tsx
<Link href="/reports"
      onMouseEnter={() => router.prefetch("/reports")}
      onFocus={() => router.prefetch("/reports")} />

Hover plus focus is the default: enough lead time to matter, high enough accuracy not to be wasteful, and it covers keyboard users.

For known funnels, prefetch the next step as soon as the user enters the current one — the checkout bundle should already be there when they finish the cart.

---

#Speculation rules

html
<script type="speculationrules">
{
  "prerender": [{
    "where": { "href_matches": "/checkout" },
    "eagerness": "moderate"
  }],
  "prefetch": [{
    "where": { "href_matches": "/*" },
    "eagerness": "conservative"
  }]
}
</script>

eagerness controls the trigger: conservative on pointer-down, moderate on hover, eager immediately. Start conservative and raise it only where the conversion is genuinely predictable.

Prerendering runs the page, including its JavaScript. Consequences:

1. Analytics will record a page view for a page nobody saw. Gate on the Page Visibility API or document.prerendering.
2. Any side effect — a POST, a counter increment, a "mark as read" — executes. Never prerender a page that mutates state on load.
3. It costs real CPU and memory on the user's device. Prerender one likely destination, not ten.

---

#Respect the user's constraints

ts
const c = (navigator as any).connection;
if (c?.saveData || /2g/.test(c?.effectiveType ?? "")) return;   // do not prefetch

1. Honour Save-Data and prefers-reduced-data. Prefetching on a metered connection spends the user's money on something they may never use.
2. Skip prefetching on slow connections — the bandwidth is needed for the current page.
3. Consider battery: speculative work on a low battery is a poor trade.
4. Prefetched responses obey Cache-Control. A resource marked no-store is fetched and discarded — pure waste. Check the headers on anything you prefetch.
5. Never prefetch authenticated or personalised URLs speculatively — the response may be cached, logged or attributed to the wrong session, and it creates load nobody asked for. → Performance/caching

---

#Anti-patterns

---

#Checklist

• Every preload specifies as, and fonts include crossorigin
• Preloading is limited to render-blocking resources
• preconnect is used only for origins certain to be needed
• Future navigations use prefetch, not preload
• Prefetching is triggered by hover, focus or a known flow step
• Speculation rules start at conservative eagerness
• No page with load-time side effects is prerendered
• Prerendered pages gate analytics on visibility
• At most one or two destinations are prerendered
• Save-Data and slow connections disable prefetching
• Prefetch targets are cacheable, not no-store
• Authenticated and personalised URLs are never speculatively fetched
• The effect of prefetching on navigation timing is measured