performance-audit

name: performance-audit description: > Post-implementation performance audit covering smoothness (流畅度), rendering speed (速度), device adaptability (设备适配), memory safety (内存安全), network resilience (网络韧性), and package size (包体积) for WeChat mini-program. Uses Gen Z phone model baselines (8GB+ primary tier, not low-end). Incorporates grill-me interview methodology to stress-test performance assumptions after every implementation. Produces gated report: PASS / WARN / BLOCK. Trigger phrases: "performance audit", "流畅度", "speed check", "device adaptability", "perf audit after implementation", "run perf check", "performance gate", "perf grill".

Performance Audit

Purpose

Run after every implementation touching apps/mini-program to audit six performance dimensions using Gen Z device baselines (8GB+ RAM, 120Hz AMOLED, 5G primary tier — Android + iPhone co-equal). Incorporates grill-me methodology: a relentless one-question-at-a-time interview about the implementation's performance characteristics.

When to use this skill / When NOT to use

• Use: After any implementation touching mini-program pages, components, routes, or data flows. Before merging a PR that changes rendering, animation, networking, or bundle composition. When a screen "feels slow" and needs a structured diagnostic. As a mandatory gate before calling a mini-program change "done."
• Skip: Pure server-side changes with zero mini-program impact. Doc-only or comment-only changes. Config-only changes that don't alter runtime behavior.

Core rules

1. Six dimensions, scored 0–10 each. See references/dimensions.md for scoring rubrics.

   • 流畅度 (Smoothness), 速度 (Speed), 设备适配 (Device Adaptability)
   • 内存安全 (Memory Safety), 网络韧性 (Network Resilience), 包体积 (Package Size)

2. Primary tier first. Optimize for Gen Z primary devices (8GB+/120Hz/5G). Degradation tier (4–6GB/60Hz/4G) is fallback-only. Do not let Degradation dictate the Primary ceiling. See references/device-baselines.md.

3. Measure before judging. Run node scripts/perf-audit-collect.mjs --changed-files=... first to collect automated evidence (package sizes, anti-patterns). Then use the grill-me interview to cover gaps the automation cannot detect (device feel, network behavior, subjective smoothness).

4. Grill-me after scoring. Once dimensions are scored, run the grill-me interview (references/grill-me-checklist.md) — one question per turn — to stress-test every performance assumption the implementer made. Do not skip any dimension with score < 8.

5. Gate thresholds:

   • PASS: composite ≥ 48, no dimension < 6
   • WARN: composite ≥ 36, no dimension < 4 → ship with documented trade-off
   • BLOCK: composite < 36 or any dimension < 4 → fix before merge

6. Report format. Produce a structured report: per-dimension score + 1-line rationale, composite score, gate verdict, and ROI-ranked fix recommendations.

Quick examples

• New animated onboarding step: score Smoothness (frame drops?), Device Adaptability (120Hz vs 60Hz?), Memory (canvas usage?), then grill-me the animation choices.
• New API-dependent page: score Speed (cold start + route transition), Network Resilience (4G behavior, timeout handling), then grill-me the data-fetching strategy.
• Bundle size regression: score Package Size (gzip measurement, subpackage placement), then grill-me every new import.

Troubleshooting

• "It feels fine on my iPhone 15 Pro" → That's not an audit. Run on representative Gen Z devices (Xiaomi 13/14, OPPO Reno, vivo X series, Huawei Pura 70). If no device available, use benchmarkLevel ≤ 15 in WeChat DevTools (or test the iOS model heuristics in apps/mini-program/src/hooks/useDeviceTier.ts).
• "The dimension score is subjective" → Tie every score to references/dimensions.md rubrics. Scores must cite specific evidence (profile trace, bundle size, cold-start measurement).
• Grill-me stalls → If the implementer gives one-word answers, expand with concrete performance scenarios. "Did you test scroll with 30+ items?" → "On a Xiaomi 13 at 120Hz, with network throttled to 4G, scrolled fast?".
• BLOCK on a trivial change → Re-classify: if the change genuinely has zero performance surface area (e.g., a copy string), the audit should pass in < 30 seconds.

Review checklist

• All 6 dimensions scored with evidence cited per references/dimensions.md
• Primary tier baselines from references/device-baselines.md applied
• Grill-me interview completed for every dimension scoring < 8
• Gate verdict (PASS/WARN/BLOCK) determined by threshold rules
• WARN verdicts include documented trade-off rationale
• BLOCK verdicts include ranked fix recommendations with estimated effort
• Report references specific files, WeChat DevTools traces, or bundle measurements

Related files

• references/dimensions.md — scoring rubrics per dimension
• references/device-baselines.md — Gen Z primary tier device data
• references/grill-me-checklist.md — performance stress-test interview questions
• docs/reference/perf.md — canonical performance strategy
• apps/mini-program/docs/DEVICE_QA_CHECKLIST.md — device QA tiers
• apps/mini-program/scripts/check-package-size.mjs — package size measurement
• scripts/perf-audit-collect.mjs — automated evidence collection (run before grill-me)
• .agents/skills/grill-me/SKILL.md, .agents/skills/frontend-performance-and-loading/SKILL.md, .agents/skills/mini-program-frontend-excellence/SKILL.md