Performance Optimization

Make the interface feel instant and measure honestly: find where users wait, decide per wait whether to eliminate it (optimistic UI, preloading, batching) or mask it (skeletons, progressive loading), and implement the fix. The prime directive: perceived latency outranks actual latency — a 3-second operation that responds instantly to the click beats a 1-second one that freezes — but never fake what you can't deliver (no invented progress numbers, no optimistic UI where rollback would burn the user). Deliver an audit ranked by user impact plus working code.

When to use / when not to

Use for: sluggish-feeling interactions, optimistic UI, Core Web Vitals remediation, preloading/anticipatory loading, request batching/dedup, layout-shift fixes, perceived-performance strategy.

Hand off instead when the real need is:

• Choosing/designing the loading indicators themselves → skills/frontend-design/loading-states
• Animation jank from non-compositor properties → skills/frontend-design/interaction-physics (motion spec) or skills/frontend-design/design-engineer-mindset (rendering pipeline, layout thrashing)
• The failure/rollback messaging your optimistic UI needs → skills/frontend-design/error-handling-recovery
• Backend/query latency itself — flag it, but that's outside this skill's frontend scope

Step 0 — Inspect the codebase, then ask only what's left

1. Find the data layer: raw fetch/axios vs a data library (TanStack Query, SWR, Apollo, RTK Query). This decides implementation strategy — libraries get optimistic updates and dedup through their APIs; hand-rolled fetching may need those primitives added.
2. Find waiting patterns: mutations that disable the UI until the server responds, serial await chains that could parallelize, duplicate fetches of the same resource, effects re-fetching on every render.
3. Find layout-shift sources: images without dimensions/aspect-ratio, content injected above existing content, fonts swapping without fallback metrics, ad/embed slots without reserved space.
4. Look for existing measurement: web-vitals package, Lighthouse CI, RUM/analytics. If the project runs in this environment, get real numbers (Lighthouse or a profile) before and after — never quote scores you didn't measure.
5. Check bundle signals when load speed is the complaint: build output sizes, missing route-level code splitting, heavyweight dependencies imported globally.

Ask the user (one batch, only what inspection didn't answer): which flows feel slow (or is it a Lighthouse/CWV score problem?), and any field data available (CrUX, analytics). If the complaint is already specific ("saving a task feels laggy"), don't ask — profile that flow and proceed.

Workflow

1. Classify every wait on the latency spectrum

2. Decide optimistic vs pessimistic per mutation — by failure cost

• Optimistic (update UI immediately, sync in background, roll back on failure): low failure cost + high likelihood of success + easy reversal. Likes, toggles, reorderings, adding items to a personal list, renames.
• Pessimistic (wait for server confirmation, show inline pending state): irreversible or high-stakes outcomes — payments, sends (email/message that actually dispatches), deletes without undo, anything with server-side validation likely to reject.
• Middle path: pessimistic commit with undo (delete now, offer "Undo" for 5–10s before the server call) gives instant feel with safety.
• Every optimistic update requires: a state snapshot before mutating, exact rollback on failure, and a user-visible failure notice (wire via skills/frontend-design/error-handling-recovery). Optimistic UI without rollback is a data-integrity bug, not a performance win.

3. Eliminate waits before masking them

In priority order:

1. Parallelize/batch requests — replace serial await chains with Promise.all or batched endpoints; deduplicate in-flight requests (data libraries do this free).
2. Preload anticipatorily — prefetch route/data on hover or near-viewport (use the framework's built-in prefetch when it exists).
3. Cache and reuse — don't refetch unchanged data on navigation.
4. Then mask what remains: skeletons for known layouts, blur-up for images, progressive enhancement for secondary content, staggered entrances for lists (cap total stagger ~300–400ms). Implementations: references/patterns.md.

4. Fix measured vitals against the real thresholds

Core Web Vitals "good" thresholds (75th percentile of field data):

• LCP ≤ 2.5s — largest content render. Fixes: prioritize/preload the LCP image, server-render critical content, cut render-blocking resources, code-split.
• INP ≤ 200ms — interaction responsiveness (replaced FID in March 2024). Fixes: break up long main-thread tasks, defer non-urgent work, avoid synchronous layout reads in handlers.
• CLS ≤ 0.1 — layout stability. Fixes: dimension all media, reserve space for async content (skeletons double here), font-display strategy with fallback metrics.

Do not invent numbers: report only measured values, label lab vs field, and if you can't measure in this environment, deliver the fixes plus instructions for the user to verify.

5. Verify

Re-measure the same way you measured before (same tool, same page, throttling noted). For optimistic flows, force a failure and confirm exact rollback + notice. For CLS fixes, reload with network throttling and watch for shift.

Required output format

Deliver both artifacts:

1. The code — implemented optimistic mutations (via the project's data library when present), parallelized/batched/deduplicated requests, masking treatments, and vitals fixes.

2. Performance Audit (markdown):

## Summary
One paragraph: what felt/measured slow, root causes, what changed.

## Measurements
| Metric | Before | After | Threshold | Source (lab/field, tool) |
| LCP | 4.1s | 2.2s | ≤2.5s | Lighthouse, throttled |
| INP | — | — | ≤200ms | not measurable here — verify via [tool] |
| CLS | 0.34 | 0.05 | ≤0.1 | Lighthouse |
(only rows you actually measured; never invented)

## Findings & fixes (ranked by user impact)
| # | Finding | Evidence | Fix | Status |
| 1 | Task save blocks UI ~800ms | network tab | optimistic update + rollback | implemented |
| 2 | 4 serial awaits on dashboard load | code | Promise.all | implemented |
| 3 | Hero image no dimensions → CLS | measured 0.34 | aspect-ratio + preload | implemented |

## Mutation strategy table
| Mutation | Strategy | Why | Rollback/undo |
| Like toggle | optimistic | trivial reversal | snapshot restore |
| Checkout | pessimistic | irreversible charge | n/a — inline pending |

## Not done / recommended next
[Backend latency, infra, CDN items outside frontend scope]

Quality bar (check before delivering)

• Every claimed number was actually measured; lab vs field labeled; unmeasurable metrics marked as such
• Every optimistic update has snapshot, exact rollback, and user-visible failure notice
• No optimistic treatment on payments, dispatched sends, or undo-less destructive actions
• No serial await chains left where requests are independent
• All images/embeds in touched code have reserved dimensions
• No fabricated progress percentages or time estimates shown to users
• Masking respects prefers-reduced-motion; staggering capped
• Fixes ranked by user impact, not by ease of implementation

Hard don'ts: don't add a data-fetching library solely for one optimization without flagging the tradeoff; don't chase a Lighthouse score with tricks that hurt real UX (lazy-loading the LCP image, skeleton-flashing instant content); don't remove existing loading feedback just to "feel faster".

Integration

• skills/frontend-design/loading-states → owns which indicator shows for waits this skill can't eliminate; run them together — space reservation is shared CLS work.
• skills/frontend-design/error-handling-recovery → supplies the rollback failure messaging every optimistic mutation needs.
• skills/frontend-design/design-engineer-mindset → owns rendering-pipeline depth (layout thrashing, frame budget) when INP/jank fixes need main-thread work.
• skills/frontend-design/interaction-physics → its compositor-only animation rule prevents the jank this skill would otherwise have to fix.

References

• references/patterns.md — optimistic create/delete/toggle with rollback, data-library notes, skeleton/blur-up/progressive/staggered masking, anticipatory preloading, request batching and dedup code, and CWV measurement guidance. Read it when implementing the fixes.