Qq Audit Performance

Part of Claude Code operator skills: a collection of skills for running a real Claude Code setup.

Audit performance with expert personas

Audit performance by running 22 expert personas in serial, fixing critical issues per framework before verifying.

22 performance frameworks, each loaded as a deep expert persona into its own agent context. Serial execution: one framework → fix critical issues → verify → next. Each agent thinks like a performance engineer with 20 years of specialization in that specific domain.

Expert lens: Is it fast? Would a performance engineer approve?

Modes

This skill responds to three modes based on the args passed. Interpret intent loosely: fuzzy matching, not exact phrasing.

Mode 1: Full serial audit (no args or "run all")

Triggers: /qq-audit-performance, "run all", "full audit", "everything", no args at all.

Mode 2: Single framework (framework name + optional scope)

Triggers: /qq-audit-performance core web vitals, "just run lighthouse on the homepage", "caching on the API routes", etc.

Match the framework name fuzzily: "cwv", "web vitals", "vitals" should all match Core Web Vitals. "lighthouse", "lh" should match Lighthouse Performance Audit. "images", "image opt" should match Image Optimization Pipeline. If ambiguous, show the 2-3 closest matches and ask.

Mode 3: List frameworks (help/list/discovery)

Triggers: "list", "what do you have", "what frameworks", "help", "which ones", "show me", etc.

Show the framework table from the Framework inventory section below, then ask which one(s) to run.

Smart interview (runs before any audit)

DO NOT ask dumb questions. Before asking anything, gather what you already know:

1. Check conversation context: What product are we working on? What files have been discussed? What has the user been complaining about?
2. Check project CLAUDE.md: Product description, tech stack, target audience.
3. Check recent session state: What was just built or changed?

Pre-fill and present assumptions:

"Here's what I know going in:

• Product: [name] ([description from context])
• Hosting: [Vercel/AWS/self-hosted, inferred from config files]
• Performance baseline: [from previous audits, Lighthouse scores, or conversation]
• Known pain points: [what the user has mentioned, such as slow pages, large bundles, API latency]
• Scope: [full app / specific routes / specific infrastructure layer]

Anything wrong or missing?"

Use AskUserQuestion with multiple choice ONLY for genuine gaps. For example, if you truly can't tell whether it's SSR or SPA, ask. If you can infer it, state the inference.

Interview output becomes the audit context: passed to every framework agent so they audit with purpose, not generically.

Execution flow (full audit)

Phase 1: Context gathering

1. Run smart interview (above)
2. Read the product's key config files (next.config, vercel.json, package.json) to understand deployment
3. Run npx next build or equivalent to establish baseline build metrics

Phase 2: Serial framework execution

For each framework (in order 1-22):

1. Spawn one agent with these instructions:

   • "Read the framework subskill file at ~/.agents/skills/qq-audit-performance/frameworks/[NN-name].md"
   • "You ARE the expert described in that document. Think like them, not like a generalist."
   • "Audit [product name] focusing on [scope]. Context: [interview output]"
   • "Previous frameworks found these issues: [cumulative findings list]. Do NOT re-report duplicates."
   • "Score 1-10, list findings with file:line references, fix critical issues (score < 7), write report."

2. Collect the agent's output:

   • Score for this framework
   • Findings (new, not duplicates)
   • Fixes applied
   • Items flagged for downstream frameworks

3. Verify build if fixes were applied (npx next build or equivalent)

4. Update cumulative state:

   • Add findings to the dedup list
   • Update the running score card
   • Note cross-framework flags for upcoming frameworks

5. Move to next framework

Phase 3: Synthesis

1. Run framework #22 (Resource Prioritization) as the final optimization check
2. Compile consolidated report:
   • Per-framework scores (1-10 each)
   • Overall weighted average
   • All findings grouped by severity
   • Fixes applied (with file:line)
   • Remaining items for future rounds
3. Compare to previous audit rounds if they exist

Phase 4: Report

Save to project's data directory:

• data/audit-performance-[date].md: full report
• data/audit-performance-[date]-summary.md: scores + critical findings only

Framework inventory

Key principles

• Serial, not parallel: 70% of findings duplicate across frameworks. Serial means each round finds genuinely new issues after fixes.
• Fix before moving on: don't accumulate a findings list. Fix each framework's criticals before the next audit.
• Expert persona, not checklist: each agent IS the specialist. They reason from principles, not rules.
• Hold every fix to a real quality bar: is this real data? Does it prevent errors? Is the complexity earned?
• Measure, don't guess: performance claims require numbers. Run builds, check sizes, time responses.
• Multi-round: after all 22 frameworks, run the full cycle again. Scores increase each round until plateau.
• Dedup across frameworks: each agent receives cumulative findings so they don't re-report known issues.