メニュー
Performance audit: bundle size, lazy loading, data fetching, caching, N+1 queries, image optimization. Native mode checks memory, I/O, launch weight, energy.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Periodic repository cleanup sweep (Tierward repo variant) - finds orphaned/merged git branches, orphaned worktrees and unregistered worktree directories, stale `.claude/session/*.md` files, stale `.claude/initiatives/*.md` files, and (target:full / target:reviews) orphaned documentation and stale `docs/reviews/` snapshots (archive-only moves, never deletion). Reports every finding with evidence, then executes only the branches/worktrees/session-files the user explicitly confirms via a single AskUserQuestion batch. Never deletes docs/*.md itself - those are always report-only, evidence-based, human-decided. Distinct from the Phase 8 cleanup gate (which tears down one specific block at closure) - this is the periodic sweep that catches what block-by-block cleanup misses.
Unified accessibility audit - axe-core WCAG 2.2 scan, APCA contrast measurement, static a11y checks (aria-label, tabindex, form labels, focus visibility, onClick on non-interactive). Static and live modes.
Responsive audit: test pages at 375/768/1024px breakpoints via Playwright. Checks overflow, tap targets, sidebar collapse, text reflow, WCAG 1.4.4 zoom.
Security audit: auth/authz on API routes, input validation, row-level access control (RLS in Postgres, ORM scopes or app-level guards elsewhere), response shape review, secret exposure, HTTP headers. Native mode checks entitlements and Keychain. MCP-aware (v1.20+): when `mcp-nvd` server is wired, Step 3c queries live CVE data instead of static `npm audit` / `pip-audit` snapshots; falls back to local audit commands when MCP unreachable.
Database audit: schema quality, index coverage, row-level access-control completeness, FK cascades, query patterns. Runs live SQL verification (PostgreSQL instance in PATTERNS.md; other engines verify the equivalent guard). Migration file safety → /migration-audit.
Code quality audit: detect cross-module coupling, N+1 queries, dead exports, antipatterns, over-large components. Cross-checks against refactoring backlog.
Replace these placeholders:
[FRAMEWORK]- e.g.Next.js App Router,Remix,SvelteKit,plain React[SITEMAP_OR_ROUTE_LIST]- e.g.docs/sitemap.md[API_ROUTES_PATH]- e.g.routes/,src/handlers/,api/[BUNDLE_TOOL]- e.g.@next/bundle-analyzer,vite-bundle-visualizer,webpack-bundle-analyzer[PERF_TOOL]- profiling tool for native/backend stacks - e.g.Instruments(Swift),Android Profiler(Kotlin),perf/flamegraph(Rust/Go),cProfile(Python),dotnet-trace(.NET). If not configured, the skill runs static analysis only and recommends profiling setup.[PROFILER_COMMAND]- command to generate a profile - e.g.xcrun xctrace record --template 'Time Profiler',go tool pprof,cargo flamegraph,python -m cProfile -o profile.out. If not configured, the skill skips the profiling step and proceeds with static analysis.- Note: if this is a public-facing app, remove the "Out of scope" restriction on Core Web Vitals
Before any other step: read CLAUDE.md and check the Framework and Language fields.
N/A - native app and NOT N/A - no web frontend): proceed to Step 1 - Web Performance Audit (Steps 1–5).N/A - native app or N/A - no web frontend): skip Steps 1–5 and proceed directly to Step 6 - Native/Backend Performance Audit.In scope: bundle composition, server/client component boundaries, dynamic rendering triggers, lazy loading, data fetching efficiency (caching, parallelism, N+1), re-render patterns, image optimization. For native stacks: algorithmic complexity, stack-specific hot-path patterns, memory management, launch weight, energy impact, binary size.
Out of scope: SEO, CWV as ranking signals, robots.txt, sitemap.xml, OG tags, Lighthouse site audit, accessibility (→ /ui-audit), DB schema (→ /skill-db).
Default mode: audit (report only, no code changes). mode:apply makes focused non-breaking fixes.
CWV reference (informational only - not primary deliverable for this internal app):
| Metric | Good | Needs work | Poor | Primary cause |
|---|---|---|---|---|
| LCP | ≤ 2.5s | 2.5–4s | > 4s | Slow server response, render-blocking resources |
| CLS | ≤ 0.1 | 0.1–0.25 | > 0.25 | Unsized images, dynamic content above fold |
| INP | ≤ 200ms | 200–500ms | > 500ms | Blocking JS on event handlers |
Source: web.dev/articles/vitals. Use these thresholds only as triage anchors - not as primary deliverables.
Parse $ARGUMENTS for target: and mode: tokens.
Operating modes:
| Mode | Behavior |
|---|---|
mode:audit (default) | Report only - no code changes |
mode:apply | Apply focused, non-breaking fixes (lazy loading wrappers, Promise.all parallelization). Describe each change before applying. |
Target resolution:
| Pattern | Meaning |
|---|---|
target:page:/dashboard | Focus on the dashboard and its component tree (example) |
| No argument | Full audit - ALL page files + layout files + lib files + API routes from sitemap.md. Maximum depth. |
STRICT PARSING - mandatory: derive target ONLY from the explicit text in $ARGUMENTS. Do NOT infer target from conversation context, recent work, active block names, or project memory. If $ARGUMENTS contains no target: token → full audit across ALL files from sitemap.md at maximum depth. When a target IS provided → act with maximum depth and completeness on that specific scope only.
Target scoping - comparative vs independent checks (mandatory): a subset of checks evaluate a project-wide property (build configuration, client bundle composition, launch entry point, release artifact); the rest evaluate one file or route in isolation.
target: filter does NOT apply to them: P1, P2, P3 (web) and NR1, NR4 (native). They MUST run against the full project even when a target is given. The client bundle, build configuration, entry point, and release artifact are global - narrowing them to the target subset hides cross-cutting exposure (a heavy package bundled by a non-targeted page still ships to every user).target: filter. Layout/shell files stay in scope regardless of target (Step 1).target: is given: apply the filter to the independent checks only; run the comparative checks across the full project and state in the report that they ran full-project by design.Announce: Running perf-audit - scope: [FULL | target: <resolved>] - mode: [audit | apply]
Apply the target filter to the file list in Step 1.
next.config.ts, vite.config.ts, nuxt.config.ts, webpack.config.js)Read docs/refactoring-backlog.md - note existing PERF- entries to avoid duplicates.
Launch a single Explore subagent (model: haiku) with all page, component, and lib files from Step 1:
"Read ${CLAUDE_SKILL_DIR}/PATTERNS.md for framework-specific grep patterns. Run all 9 checks below. For each: state total match count, list every match as file:line - excerpt, and state PASS or FAIL.
CHECK B1 - Unnecessary client-side rendering scope Find components or pages marked for client-side rendering that do not actually use browser APIs (event handlers, state, refs, browser-only hooks). See PATTERNS.md → B1 for client-side directives per framework.
CHECK B2 - Heavy libraries in client bundle Grep for imports of known-heavy libraries in client-rendered files. Flag any imports of libraries that do NOT need browser APIs and could run server-side:
CHECK B3 - Client-side data fetching that defeats server rendering Find client-rendered components that fetch data in lifecycle hooks (see PATTERNS.md → B3) instead of using server-side data loading. Flag: data fetching in client lifecycle hooks - these defeat server rendering and add a loading waterfall. Use server-side data loading with streaming/suspense, or a client-side cache library with proper revalidation.
CHECK B4 - Unstable callback references defeating memoization Find inline arrow functions passed as props to memoized child components. Inline functions create new references on every parent render, defeating memoization. Flag: each case. Skip native HTML element event handlers - frameworks batch these internally.
CHECK B5 - Sequential await waterfall
Pattern: two or more consecutive await calls for independent data sources in the same async function.
Grep in server-rendered files or API routes: lines matching const .* = await that appear consecutively (within 5 lines of each other) without the first result being an input to the second call.
Flag: each pair of sequential awaits that could be parallelised with Promise.all (or equivalent).
Example of violation: const a = await getA(); const b = await getB(); where b doesn't depend on a.
CHECK B6 - Missing caching on repeated server-side queries Find data fetches that run on every request without caching in server-rendered components. Grep in server-rendered files for database query calls - check if they use any caching mechanism (framework cache, memoization, or explicit cache headers). Flag: uncached queries in layout-level or frequently-visited components (e.g. navigation data, user profile). These execute on every page visit.
CHECK B7 - Images without explicit dimensions (CLS risk)
Find image elements (framework image component or native <img>) without width/height or fill/cover constraints.
Flag: each unsized image. Without dimensions, the browser cannot reserve layout space - content shifts when the image loads, causing CLS violations.
CHECK B8 - Accidental dynamic rendering triggers Find patterns that force dynamic rendering at the layout level, opting entire route subtrees out of static generation:
CHECK B9 - Missing lazy loading for heavy client components Find imports of heavy libraries (rich text editors, chart libraries, complex UI components) that are statically imported without lazy loading. Flag: any file where a heavy component is eagerly imported without using the framework's dynamic/lazy import mechanism (see PATTERNS.md → B9)."
Read the framework configuration file.
P1 - Bundle analyzer availability Check if a bundle analyzer is configured or available for the project's build tool (see PATTERNS.md → P1 for tool names per build system). If no analyzer is configured or documented, flag as Medium - developers cannot easily audit bundle composition.
P2 - Server-only package exclusion Check if the framework configuration excludes heavy server-only packages from the client bundle. Identify packages in the project that should never be client-bundled (e.g. PDF processors, XLSX generators, database drivers, file system utilities). Flag: any heavy server-only package not explicitly excluded from client bundling.
P3 - Tree-shaking optimization for large libraries Check if the build configuration optimizes imports for large libraries with many exports where only a subset is used (e.g. icon libraries with hundreds of icons, utility libraries like lodash). Note: some frameworks automatically optimize certain packages (check the framework docs). If a library is already auto-optimized by the build tool, note as PASS. Flag: any package with 100+ exports where only a subset is used, not covered by the build tool's tree-shaking or import optimization.
"Run these 3 checks. Adapt grep patterns to the project's database client or ORM:
CHECK Q1 - Unbounded queries (no limit on large-growth tables) Flag: each collection query without pagination bounds. Exception: export routes that intentionally fetch all for CSV/XLSX - verify from context.
CHECK Q2 - Select * (over-fetching columns) Grep for select-all patterns in route handlers (see PATTERNS.md → Q2 for ORM-specific patterns). Flag: each match. Fetching all columns is a performance and security risk - columns with large values are sent over the wire unnecessarily.
CHECK Q3 - N+1 patterns
Pattern A: database query calls inside .map(, for, or forEach loops.
Pattern B: sequential await with a DB client call inside a loop body.
Flag: each match. N+1 on a list endpoint means N DB queries for an N-item list - use batch queries or embedded/eager loading instead."
Generate the report using the template in ${CLAUDE_SKILL_DIR}/REPORT.md (Web Audit section). Apply the severity guide and backlog writing rules from the same file.
mode:audit (default): do NOT make any code changes - report only. After producing the report, ask: "Should I implement the High/Critical priority optimisations?"mode:apply: apply only the fixes listed in Quick wins (isolated, non-breaking). Describe each change before writing it. Do not apply Strategic refactors without explicit user confirmation. Do NOT ask the closing question - the user already expressed intent via mode:apply.This path runs for non-web projects only. Web projects use Steps 1–5 above.
Primary tool: [PERF_TOOL]
Profile command: [PROFILER_COMMAND]
Run the profiler on the main execution path. Identify the top 5 hotspots by CPU time and the top 5 by memory allocation. If profiling data is not available, proceed with static analysis and recommend running the profiler.
Launch a single Explore subagent (model: haiku) with all source files from the project:
"Run these 4 checks on the provided source files. For each: state total match count, list every match as file:line - excerpt, and state PASS or FAIL.
CHECK NP1 - Nested iteration on collections Grep for nested loops (for/while/map/forEach inside another loop body) operating on collections or arrays. Flag: O(n²) or worse complexity. Exclude small fixed-size iterations (< 10 items known at compile time).
CHECK NP2 - Memory allocation in hot paths Flag:
CHECK NP3 - I/O bottleneck patterns Flag:
CHECK NP4 - Concurrency inefficiency Flag:
Read the 10 largest source files by line count. Apply the checks relevant to the project language using grep patterns from PATTERNS.md → Step 6b (organized by language). Run the patterns, then read flagged files for context.
CHECK NR1 - Launch / startup weight Identify the application entry point (see PATTERNS.md → NR1 for per-language markers). Grep for heavy operations in the entry point: database initialization, network calls, large file reads, complex object graph construction. Flag any operation that could be deferred (lazy initialization) or moved to a background thread/task.
CHECK NR2 - Memory management patterns Grep for per-language memory antipatterns from PATTERNS.md → NR2. Key categories: missing autorelease pools in batch loops, unbounded caches, leaked references, pre-allocation misses. Also flag large static/global collections that persist for the process lifetime.
CHECK NR3 - Energy and background patterns (mobile stacks only) Grep for per-language energy antipatterns from PATTERNS.md → NR3. Key categories: tight timers without tolerance, high-frequency location updates, excessive wake-ups. Flag any background polling pattern that could use push notifications or event-driven updates instead.
CHECK NR4 - Binary / artifact size
Generate the report using the template in ${CLAUDE_SKILL_DIR}/REPORT.md (Native Audit section). Apply the severity guide and backlog writing rules from the same file.
mode:audit (default): report only. After report, ask: "Should I implement the High/Critical priority optimisations?"mode:apply: apply only Quick wins. Describe each change before writing.