Optimize Next.js Fluidity

Turn an unknown Next.js repo into a measured, ordered, executed optimization pass — without ever recommending something the installed version does not support.

This skill profiles the repo, gates every practice against what that install actually accepts, fans out parallel audit agents, writes a dependency-ordered plan as one markdown file per task under nextjs-enhancement/, then executes the safe tasks and verifies each one. One-way doors are prepared for a human, never auto-applied.

The knowledge baseline is Next.js 16.3.0 / React 19.2, verified 2026-08-05. The skill does not assume the target repo is on that version — that is the entire point.

When to use

• "audit this Next.js app for performance", "make the app feel faster / more fluid"
• "apply Next.js best practices to this repo", "find and fix our Core Web Vitals problems"
• "we're on an older Next — what should we adopt, and what should we not?"
• "plan and execute a performance pass, one task per file, so I can review the trail"
• After a Next.js major upgrade, to find dead flags and newly available features

Do not use for:

• A single known bug. Fix it inline; this skill's recon and gating overhead pays back across many routes and domains, not one component.
• Generic code review. Use run-review. This skill is performance/fluidity-specific.
• Visual/CSS bugs. Use audit-ui-ux — that one drives a vision-first evaluation.
• Non-Next.js projects, or Pages Router-only apps. The corpus is App Router-shaped. Pages Router surfaces appear only as migration notes.

Two non-negotiables

1. Nothing is recommended without a capability probe

Version arithmetic is not a gate. The skill reads the installed package — node_modules/next/package.json and the bundled config-schema.js — and asks whether each config key actually exists on this install. Three failures this prevents, all observed on a real 16.2.9 repo during design:

• partialPrefetching (16.3-only) probes absent → withheld entirely, even though its documented prerequisite cacheComponents was satisfied.
• experimental.viewTransition is recorded as removed in the 16 line, but probes present and is deliberately set → no removal task, because deleting it would have been a breaking change. At most a note to re-check on upgrade.
• staleTimes probes present and is set → the finding is justified by its stability tier (experimental, production-discouraged), never by absence.

Full mechanism and the withhold-list discipline: references/gating/capability-probe.md. The graveyard of removed/renamed surfaces and every version floor: references/gating/version-matrix.md.

2. Execution is autonomous, but one-way doors are not

There is no mid-run approval prompt. Safety comes from hard rails instead: a task whose reversibility is migration-required is written with Status: blocked-needs-human and its pre-flight checklist, and is reported as prepared work — never executed. Fixes run on a branch, one commit per task, and a task is done only after its own verification command passes; a failure is reverted and reported with output.

The complete contract — dirty-tree refusal, side-effect bans, the severity/reversibility auto-apply matrix, and the stop conditions — is references/workflow/safety-rails.md.

The workflow — six phases

Phases 1, 2, 4, and 6 are orchestrator-personal. Delegating judgement produces a plan nobody owns.

Phase 1 — Recon (personal)

python3 scripts/recon.py <repo-root>              # profile → markdown
python3 scripts/probe-capabilities.py <repo-root> # the gate primitive

Both are read-only. Paste their output into nextjs-enhancement/00-recon.md, then add the judgement calls only a reader can make: archetype, baseline availability, and an explicit "existing good practice — do not re-propose" list. Shape and rules: references/artifact/recon-report-template.md.

Phase 2 — Applicability gating (personal)

Evaluate every gate row in each references/detect/*.md against the recon facts. Produce nextjs-enhancement/01-applicability.md: one verdict per domain (APPLICABLE · APPLICABLE-WITH-REMOVAL · BLOCKED-PARTIAL · APPLICABLE-CUSTOM · NOT-APPLICABLE) plus the withhold list — features this skill knows but must not recommend here. A NOT-APPLICABLE domain gets zero agents. Verdict computation rules and worked rows: references/artifact/applicability-template.md.

Then scaffold the artifact tree and plant the format authority:

python3 scripts/scaffold-plan.py <repo-root> --apply \
  --domains <applicable,domains> \
  --readme references/artifact/artifact-readme.md

Dry-run is the default; --apply creates directories and README.md only — never content. The plantable body lives in references/artifact/artifact-readme.md.

Phase 3 — Audit wave (parallel agents)

One nextjs-perf-auditor per applicable domain, dispatched in one message. These are read-only Grep/Read agents writing to disjoint folders — no shared lock, so full parallelism is correct here (unlike browser-driven audits, which must stagger).

Each agent gets: its domain's references/detect/<domain>.md, its applicability verdict, a short repo-profile excerpt, and a write scope of findings/<domain>/ only. Brief template: references/subagent-briefs/audit-brief-template.md. Caps, gates, and partitioning: references/subagent-briefs/dispatch-rules.md.

Finding shape — literal file:line, the matched text, the command that found it, and six cleared false-positive checks: references/artifact/finding-template.md.

Zero findings is a valid result. A clean domain writes nothing.

Phase 4 — Synthesize the plan (personal)

Read every finding. Cluster them into tasks — twelve call sites of one deprecated prop is one task against the shared wrapper, not twelve. Order by dependency, not by discovery: references/gating/priority-matrix.md supplies both the hard dependency graph and the per-archetype priority tiers; references/gating/composition-recipe.md supplies the 14-step blueprint each task maps onto.

Write one file per task (references/artifact/task-template.md) plus the index (references/artifact/index-template.md). Every task carries evidence, a recipe pointer, the exact change, a verification command, a rollback, and its reversibility.

Set reversibility from references/gating/lockin-reversibility.md. Check every task against references/gating/conflicts.md — adding a transition to a route that suspends, or a cookie-read to a cached root, produces changes that look right and behave wrong. Cost direction for anything touching caching, images, prefetch, or build machines: references/gating/cost-model.md. SEO obligations for anything touching rendering, routing, or metadata: references/gating/seo-obligations.md.

Phase 5 — Fix waves (parallel agents)

Cluster auto-apply tasks by shared target files so no two agents edit the same file. Dispatch in dependency-respecting waves: prerequisites and cleanups first, architectural work second, work that assumes the architecture third. Gate between waves by reading task statuses. Brief: references/subagent-briefs/fix-brief-template.md.

Each agent applies only its task's stated change, runs the verification, commits one task per commit, and on failure reverts and records the output.

Phase 6 — Verify and reconcile (personal)

python3 scripts/status-report.py <repo-root>/nextjs-enhancement

Read-only; exits non-zero on integrity problems — a done task whose dependency is not done, a migration-required task marked applied, a done task with no verification section. Then write nextjs-enhancement/02-completion-report.md (references/artifact/completion-report-template.md): applied and verified · prepared but blocked on a human · failed and reverted · withheld and why.

Verification vocabulary and the honesty rules for claiming a rung: references/workflow/verification-playbook.md.

The 14 domains

Each has a detect file (gate table, detection commands, severity rubric, false-positive filters, pitfall signatures) and a fix file (version-annotated recipes with Why / Verify / Lock-in / Rollback). Full set: references/detect/*.md and references/fix/*.md.

Domains are library-optional where the ecosystem is: a repo with custom theming or custom i18n gets APPLICABLE-CUSTOM, and the fix files carry custom-implementation variants. Proposing a library migration for a working custom implementation is a false positive.

Writing new domain references

The two contracts every domain file obeys: references/workflow/detect-file-contract.md and references/workflow/fix-file-contract.md.

Footguns

1. Dispatching before gating. Agents for NOT-APPLICABLE domains produce noise and waste the wave. Phase 2 exists to prevent it.
2. Version arithmetic instead of the probe. 16.2.9 ≥ 16 is true and useless — the feature you want may have shipped in 16.3.
3. Blind-deleting a graveyard flag. If the install still accepts it, removal is a breaking change. Probe first.
4. Treating availability as endorsement. staleTimes exists on most installs and is still production-discouraged.
5. One task per grep hit. Cluster to the shared wrapper, or the fix wave will fight itself over one file.
6. Auto-applying a one-way door. cacheComponents, URL policy, canonical changes — prepared, never executed.
7. Marking done without running the verification. Reading the diff is not evidence.
8. Flagging deliberate divergence. Tuned config with an explanatory comment is informational. See references/workflow/false-positives.md.
9. Counting comments as usage, or counting test files, RSS routes, and ImageResponse contexts against page rules.
10. Animating before the destination is cache-hot — the "smooth but wrong" trap.
11. Recommending a library to a working custom implementation.
12. Re-proposing what the repo already does correctly. Recon lists it; honour the list.

Scripts

All take the target repo (or plan root) as an explicit argument, so they work from any working directory.

Bottom line

Profile the repo, gate every practice against the installed package, audit in parallel, plan one file per task in dependency order, execute what is safely reversible, prepare what is not, and verify everything you claim. The nextjs-enhancement/ folder is the deliverable — it outlives the session and tells the next reader exactly what was decided and why.