Next Cache Components Optimizer

next-cache-components-optimizer

Set up an agentic optimization loop that drives a Next.js route from "not instant" to "instant" and keeps it there. The loop is test-driven: encode the goal as a failing @next/playwright instant() test, work it to green, and ship the test as the regression guard. Run it once per target route. Work the phases P → G in order; each ends in a gate. Fix recipes live in two lazily-read references — reference/patterns.md (before→after for each blocker type) and reference/real-app-patterns.md (parallel routes, auth gates, the empty-shell and responsive-skeleton failure modes). Read one only when its phase points there.

What is invariant, and what is yours

One thing here is fixed. The rest is yours. Read this before treating any command, platform, or env var below as a requirement.

• Invariant: the verification loop. Maximizing the shell is worthless unless you can prove it. The proof is an automated check: under a lock that gates dynamic data, the static shell still commits. RED shows the gap, GREEN shows it closed, the test ships as the regression guard. It must run on a production-like build and must not be able to pass vacuously. Stand the loop up once; every later optimization is then verifiable by construction. The loop is the deliverable, not any one route.
• The mechanism: @next/playwright instant(). This skill uses instant() as a ruler, not a stopwatch (phase A). It comes from @next/playwright (installed alongside @playwright/test, on the same release line as next), so it isn't tied to any host. Keep it. Timing a navigation by hand is too flaky to trust, and is the failure mode this skill exists to prevent.
• Yours: the rig. How you build, deploy, authenticate, configure Playwright, and loop belongs to your stack, not to this skill. A local next build && next start, a CI/staging container, and a per-push preview deploy are equally valid rigs; the verdict comes from the build, never the platform. Phase 0 maps the invariant onto your repo. Read every platform name, env-var spelling, and command below as an example to translate, not a requirement.

Two navigations, two loading states

A route reaches the user two ways, and both must be instant:

• Initial load (hard navigation) commits the route's prerendered static shell; deferred parts stream in behind their loading skeletons (Suspense fallbacks, loading.tsx).
• Client-side navigation (soft navigation) commits the destination's prefetched App Shell — the <Link> default under Partial Prefetching — re-rendering only the segments that change.

The fix patterns are identical for both; the test differs only in how the navigation is driven ("Driving the navigation in tests" below). The two shells can differ; guard the one you ship, both when both matter (reference/real-app-patterns.md).

Goal

Maximizing the static shell is the optimization objective: the most meaningful prerendered content commits immediately, and only genuinely per-request data streams in afterward. The shipped test deterministically encodes present ∧ instant; non-blank is the additional bar the workflow enforces by judgment (D1/D2/E), because an instant() pass alone is satisfied by a blank fallback={null} shell (the empty-shell failure mode, reference/real-app-patterns.md).

instant() is a ruler, not a stopwatch: assert that the shell appears under the lock; do not time it. A trustworthy verdict requires a production build (phase A).

The GREEN under the lock is the deterministic verdict; each gate keeps it trustworthy.

Reporting to the user

This loop is meant to run unattended, so it doesn't stop to ask between steps. Work the navigation the user named, finish it, and stop. What matters is how you word and present the results, not how often you interrupt. The mechanics below — the rig, RED, GREEN, the gates — are your scaffolding; the user never needs to hear those words.

• Speak their language. Describe the gap and the result in terms of what the user sees: "navigating to the dashboard waited on the charts query before anything painted; now the layout and skeletons paint instantly and the charts stream in" — not RED/GREEN, the lock, or the phase letters.
• Show, don't tell. When you report a route, drive the browser (or attach before/after screenshots) so the user watches the shell commit immediately and the data stream in, rather than reading a claim. Identical before and after means the fix did nothing — roll it back.
• Present a run as a list of results the user can click through — one line per navigation: the route, what commits instantly, and what streams in — not a transcript of the loop.
• Only surface a question for a genuine fork: a fix that would change behavior, a security-sensitive read, or a route that's dynamic by design (a per-link-prefetch candidate, not a shell to grow). A clean instant fix is not a fork — keep going. With no one to ask (an unattended run), don't block: take the safe default and note the assumption — for a cache-freshness choice, defer the read behind <Suspense> (always fresh, still instant) rather than guess a cacheLife.

The workflow

- [ ] P  PREREQS      Next.js 16.3+ with cacheComponents: true; upgrade first → below
- [ ] 0  SETUP        once per repo: discover + write instant-nav.rig.md     → rig-template.md
- [ ] A  RIG          production build with the testing API exposed          → below
- [ ] B  BASELINE     unlocked: the marker renders for the test user         → test-template.md
- [ ] C  RED          locked instant(): the shell does not commit            → test-template.md
- [ ] C-gate          VERIFY-RED: stop until the RED is trustworthy          → reference/red-test-robustness.md
- [ ] D  FIX          push each Suspense boundary down to the data it guards → reference/patterns.md
- [ ]      D1 reuse the route's existing loading UI; do not hand-build skeletons
- [ ]      D2 the shell matches the real render at every breakpoint  → reference/real-app-patterns.md
- [ ] E  PARITY       the refactor changed only whether the route is instant
- [ ] F  DIFFERENTIAL revert only the fix → RED; re-apply → GREEN            → reference/red-test-robustness.md
- [ ] G  REVIEW       PR checklist (below)

Phases B and C build the test; only the locked test from C ships.

---

P. PREREQUISITES: current Next.js with Cache Components

The workflow depends on framework capabilities that ship with current Next.js:

• Next.js 16.3+ with cacheComponents: true in next.config.ts. Without Cache Components there is no static shell to optimize.
• @next/playwright on the same release line as the project's next; it provides instant(). Verify with npm ls next @next/playwright (or the project's package manager) and align them if they differ. The matching testing API is in the next runtime, gated by the experimental.exposeTestingApiInProductionBuild config flag (phase A).

If the project does not meet these, upgrade first (npx @next/codemod upgrade automates most of it), then enable Cache Components in next.config.ts:

export default { cacheComponents: true }

Enabling the flag surfaces the blocking routes to resolve first; the next-cache-components-adoption skill drives that adoption. Reach for this optimizer once the app builds under Cache Components.

This gate is deliberate: the skill targets current Next.js, and none of the verdicts below are meaningful on older versions.

0. SETUP: discover this project's rig, once per repo

The principles in this skill are fixed; the infrastructure they run on is yours. On first use in a repository, discover how the project builds, deploys, authenticates, and tests (inspect the repository first, and ask the user only what it cannot answer), then write the answers to a committed instant-nav.rig.md. Every later run reads that file instead of rediscovering. The six questions (BUILD / EXPOSE / RUN / TEST USER / DRIFT / LOOP), the file template, and filled examples (local-only, generic CI + container, preview deploy) are in rig-template.md.

If the repo has no Playwright e2e harness yet, standing up a minimal one (@next/playwright, a config with baseURL, one authenticated path) is part of this step; the loop does not assume a pre-existing suite.

A. RIG: a production build with the testing API exposed

Stand up the rig described by instant-nav.rig.md. Two invariants hold on every platform:

1. Never measure on next dev. It does not prefetch, and its lock is unreliable for blocking routes, so a dev instant() result is not a valid RED or GREEN.

2. The measured build must expose the testing API. Otherwise instant() silently no-ops and the test passes vacuously (see reference/red-test-robustness.md). The lock-engagement proof is the phase-C RED itself: the unfixed target route is the known-blocking route, and its RED under the lock shows the lock engages on this build (C-gate); the self-validating variant in test-template.md is the in-band guarantee. Wire experimental.exposeTestingApiInProductionBuild to a condition that is true for every build you measure and never true in production:

   experimental: {
     // Use the condition your platform provides, and record it in the rig file:
     //   local:       an explicit opt-in, as below
     //   generic CI:  process.env.DEPLOY_ENV === 'staging'
     //   Vercel:      process.env.VERCEL_ENV === 'preview'
     exposeTestingApiInProductionBuild:
       process.env.EXPOSE_TESTING_API === '1',
   }
   

The rig is any production-like build that exposes the testing API: a local next build && next start, a CI/staging container, and a preview deploy are all equally valid; the verdict comes from the build, not the platform. See rig-template.md for filled examples.

For any deployed or remote build, poll the rig's LIVENESS probe to confirm the artifact contains HEAD before trusting a verdict (a stale deploy reads as a false RED or GREEN); a local next build && next start needs none. The probe mechanism is in rig-template.md (question 6).

B. BASELINE (unlocked): development scaffold, do not ship

Drive the real navigation with no instant() lock and assert that the destination's SHELL_MARKER renders as the test user: the account the e2e suite authenticates as (in CI, the CI account; locally, your e2e login fixture), with its flags, plan, role, and data. This establishes that the marker is real and reachable: not flag-gated, not redirected away, not a guessed selector. The suite runs as the test account, not the author's session; that environment drift (the rig DRIFT list) is a common source of untrustworthy REDs. Scaffold and run command: test-template.md. Delete this baseline before the PR.

C. RED (locked) + the VERIFY-RED gate

Wrap the same navigation in instant(); assert the shell commits under the lock. A RED here is the gap. This is the test that ships (test-template.md).

Prefer the self-validating variant when the route has deferred content. If the route cannot build while blocked, or a cookie/session read stays GREEN, use the RED recipes in reference/red-test-robustness.md.

C-gate: do not start optimizing until the RED is verified trustworthy. A RED that is red for the wrong reason sends you optimizing a route that was never broken.

The question that settles it: does SHELL_MARKER render without the lock, as the test user? Answer it by re-running phase B as the test user, not by adding assertions to the shipped test. The two-branch resolution (No → marker or environment bug; Yes → genuine gap, proceed to D), the full taxonomy of untrustworthy REDs, the checklist, and worked cases are in reference/red-test-robustness.md. Read it now.

---

D. FIX: push each boundary down to the data it guards

The anti-pattern: one coarse boundary. A single <Suspense> high in the tree with a page-level fallback has three costs:

• The layout UI stays out of the static shell: only a throwaway copy of it is prerendered.
• The entire subtree is replaced when the boundary resolves, which discards client state and shifts layout.
• The hand-built fallback drifts out of sync as the UI changes, because it duplicates structure that also exists in the resolved tree.

The fix: hoist the static, push the Suspense down. Render the layout UI once, synchronously, in the shell, and wrap each await in a boundary scoped to the single read it guards. Only that leaf streams; the stable ancestors are reused as-is.

Rule: if an element renders in both the fallback and the resolved tree, hoist it above the boundary.

The most common blocker: a top-level await in a layout on a fallback route

app/[locale]/(app)/[tenant]/dashboard/...
       │ generateStaticParams ✅   │ no generateStaticParams → fallback route

When any dynamic segment in the route lacks generateStaticParams, the route is a fallback route, and all params defer to request time, including the enumerated ones. A top-level await in a layout (await params, a request-time session read, an auth gate) then blocks the whole subtree out of the static shell, even when it reads a statically known param. Minimal shape: a dynamic-segment route with one segment lacking generateStaticParams, plus a top-level await in the layout above it.

The fix: defer the gate, render children

Render children unconditionally; move the top-level await into a <Suspense fallback={null}>-wrapped child. Mechanism and before→after: reference/real-app-patterns.md, "Deferring an auth gate".

Fix the page below the shell too, not only the layout. A page-level top-level await (commonly await params) blocks the same way the layout's does, so make the page sync and push its dynamic reads into a <Suspense>-wrapped leaf as well. fallback={null} is correct only when a gate renders nothing on success; for data, the fallback must be a real loading skeleton (see D1).

Every other blocker shape — cookies()/headers(), uncached fetch or database reads, searchParams, metadata, viewport, non-deterministic values (Date.now(), Math.random(), crypto.randomUUID()) — surfaces its own insight when you hit it: the build prints a https://nextjs.org/docs/messages/<slug> link. The default build output is often abbreviated and may carry no usable stack trace; add --debug-prerender for the full failing frame and to report every blocker past the first. Scope the build to the route you're on with next build --debug-build-paths "app/<route>/**" rather than rebuilding the app. Open that page and apply its recipe; don't improvise from the inline message.

The before→after recipe for each shape is in reference/patterns.md, which maps it to the insight that explains it.

A few things those per-error pages don't stress for the instant-navigation goal:

• A boundary in the root layout isn't enough for client navigations. It passes a page-load check but leaves sibling client navigations blocking; put the boundary below the lowest layout the source and destination routes share.
• Keep the LCP element (usually the main heading) out of any boundary, so it paints in the shell instead of waiting on a stream.
• A green check isn't always instant. export const instant = false opts the segment out of validation while the navigation still blocks, and a <Suspense> above the document <body> prerenders an empty shell — neither makes the route instant.

D1: reuse the route's existing loading UI; do not hand-build skeletons

Before writing any skeleton, search the repository for the loading UI that already exists for this route, in order:

1. the route's loading.tsx;
2. an exported *Skeleton colocated with the component;
3. the fallback already inside the component's own <Suspense>.

The divergence point is the lowest layout shared by the source and destination routes: a soft navigation re-renders only the segments below it, while an initial load re-runs every layout from the root. (Also called the shared boundary.) A loading.tsx above the divergence point fills only the initial-load shell; it sits above the soft-nav re-render scope. A loading.tsx at the destination segment is itself the in-tree boundary for a soft navigation into that segment and serves both. Reuse whichever boundary actually covers the navigation you are shipping; below the divergence point, loading.tsx and colocated skeletons are interchangeable for that purpose.

If a component has no skeleton, extract its loading markup into a colocated skeleton beside it. Do not author a fresh skeleton that mirrors the page layout: it duplicates structure, drifts as the page changes, and pulls the design back toward a single coarse boundary. Reusing the component's own skeleton also keeps the prefetched shell consistent with the loaded UI.

See: Streaming and loading states.

Exception: if the deferred component renders null for some users (for example, a flag-gated control), fallback={null} is correct, since a skeleton would flash and then collapse.

D2: the shell must match the real render at every breakpoint

A skeleton frozen to one breakpoint misaligns on the others. Fix it the same way: one responsive component renders both the live UI and the shell (D1 skeleton in its data slots), so the breakpoint switch happens once. Verify by re-asserting the shell marker at two widths (await page.setViewportSize({ width: 1280, height: 800 }), then { width: 390, height: 844 }), or by adding a mobile Playwright project, so this gate is as machine-checkable as the others. Detail: reference/real-app-patterns.md.

D-gate: phase D is complete when the locked test from phase C passes GREEN under the lock on the production-build rig, not when the code compiles. That GREEN is the deterministic stop for the fix loop; proceed to E.

When URL data can't be pushed down (for example, the whole page depends on params, searchParams, or the full URL), there may be no meaningful static shell to grow. Don't force one. Per-link prefetching can make the soft navigation instant, but it is outside this optimizer loop: it requires Partial Prefetching, a <Link prefetch={true}>, and cached URL-dependent content. See Optimizing prefetching and pattern 10 in reference/patterns.md for the requirements, cost trade-offs, manual prefetch caveat, and instant() test gotchas.

E. PARITY: the refactor changed only whether the route is instant

The push-down is a mechanical transform, not a redesign. Afterward the route must render the same tree, data, ordering, empty and error states, redirects, and interactions as before; the only observable difference is that the shell now commits instantly. Verify:

• Same render output. The moved awaits compute and return the same values; after the stream, the route shows the same content as the base branch for the test user.
• Side effects still fire. A deferred redirect() or notFound() still happens, at request time rather than during prerender. Confirm an unauthorized user is still redirected and a missing record still returns 404.
• Both viewports reach the real UI after the stream (D2).
• Client state survives. Because the layout UI is hoisted into the stable shell rather than swapped on resolve, open menus, scroll position, focus, and input state persist across the stream.
• Pre-existing failures stay separate. If the route errors after the change, reproduce it on the base branch. The same failure there is an environment or data problem, not an optimizer regression.

If anything other than whether the route is instant changed, reduce the refactor.

F. DIFFERENTIAL

Revert only the fix → RED; re-apply → GREEN; link both runs (reference/red-test-robustness.md). On a deployed rig, confirm each run is live (LIVENESS, phase A) before trusting its color.

G. REVIEW (PR checklist)

A green final state means nothing if the RED was never trustworthy. The test-trustworthiness items are the robustness checklist (reference/red-test-robustness.md); confirm them, then require these PR-specific items:

• Differential shown: RED without the fix, GREEN with it, runs linked.
• Parity confirmed (E): same content, redirects, and state.
• Existing loading UI reused (D1): no new page-mirroring skeleton.
• Shell matches the real render at desktop and mobile widths (D2).
• Baseline removed: only the locked test from C remains.

Stop condition for the whole workflow: the locked test from C is GREEN on the rig, the differential (F) holds, and every item above is checked. Until all three hold, you are not done.

Driving the navigation in tests

• Soft navigation → drive a real <Link> click. Initial load → use page.goto() inside instant() with the baseURL option. Do not substitute goto for a soft-nav verdict; the two shells can differ (test-template.md, reference/real-app-patterns.md).
• With parallel routes, only the slots that change re-render on a soft navigation; client-rendered navigation UI does not re-render at all. Do not chase a slot the navigation never touches (reference/real-app-patterns.md).

Files

• rig-template.md: phase 0, the six-question rig discovery, the instant-nav.rig.md template, and filled examples (local-only, generic CI, preview deploy).
• test-template.md: the shipped instant() specs for both navigation types (phase C), and the delete-before-PR baseline scaffold (phase B).
• reference/red-test-robustness.md: the C-gate and phase F. The taxonomy of untrustworthy REDs, the checklist, the differential recipe, the vacuous-pass failure mode, and worked cases.
• reference/real-app-patterns.md: parallel routes, deferring an auth gate, initial-load vs soft-navigation shells, the empty-shell failure mode, the responsive-skeleton mismatch, edge cases.

After optimization

Once the target routes are instant, check whether the app has already adopted Partial Prefetching (partialPrefetching: true, or the relevant destination still uses prefetch = 'partial' during an incremental rollout).

Make that check mechanically:

rg -n "partialPrefetching|prefetch\s*=\s*['\"]partial['\"]" --glob 'next.config.*' --glob 'app/**' --glob 'src/app/**'

If partialPrefetching: true is in config, the app is globally adopted. If only prefetch = 'partial' matches, treat those destination segments as adopted during an incremental rollout and keep checking any other target routes.

• Already adopted: for any URL-data route that stopped at the limitation above, consider a targeted <Link prefetch={true}> on the links where having that URL-specific content ready before the click is worth the per-link server work. Keep the default link behavior everywhere else so the shared App Shell remains the low-cost baseline.
• Not adopted yet: recommend next-partial-prefetching-adoption. That skill moves the app onto the better prefetching model: shared App Shell prefetches by default, fewer duplicated full-prefetch requests for visible links, a link audit for existing <Link prefetch={true}> usage, and optional per-link prefetching only where URL-specific content is worth the extra server work.