website-audit
Full website SEO audit — crawls up to 500 pages, detects business type, spawns up to 15 parallel specialist agents (8 always + 7 conditional), generates SEO Health Score (0-100) with enterprise HTML report. Completely standalone — not connected to the blog-writer pipeline. Triggers on: audit my site, website audit, full SEO check, site health, analyze my website, SEO analysis, technical SEO audit, or any URL + audit intent. Use whenever user mentions 'audit', 'check my site', 'SEO health', 'website analysis', or provides a URL asking for SEO evaluation.
XuanRanL · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add XuanRanL/loamwright-SEO-Skill --skill website-audit技能说明
name: website-audit description: "Full website SEO audit — crawls up to 500 pages, detects business type, spawns up to 15 parallel specialist agents (8 always + 7 conditional), generates SEO Health Score (0-100) with enterprise HTML report. Completely standalone — not connected to the blog-writer pipeline. Triggers on: audit my site, website audit, full SEO check, site health, analyze my website, SEO analysis, technical SEO audit, or any URL + audit intent. Use whenever user mentions 'audit', 'check my site', 'SEO health', 'website analysis', or provides a URL asking for SEO evaluation." user-invocable: true argument-hint: " [--max-pages N] [--output dir]" allowed-tools: [Read, Write, Bash, Glob, Grep, Agent, Task]
Full Website SEO Audit
Standalone website auditing tool. URL in → crawl → analyze → report out. No dependency on the blog-writer pipeline, project system, or article creation flow.
Invocation
/website-audit https://example.com
/website-audit https://example.com --max-pages 100
Process
Phase 1: INITIALIZE
- Validate URL:
python -c "from scripts._core.ssrf_guard import validate_url, SSRFError; import sys
try:
validate_url('$URL', allow_http=True); print('OK')
except SSRFError as e:
print(f'BLOCKED: {e}'); sys.exit(1)"
⚠️ validate_url() signals success by RETURNING a truthy ValidatedURL and failure by RAISING SSRFError. The old snippet here did err = validate_url(...); print(err or 'OK'), which prints the success object and can never print OK — and that inverted reading got copied into verify_backlinks.py, where it marked every valid backlink "blocked" and silently disabled link verification entirely. Never treat the return value as an error indicator.
2. Create output directory: ./Website Audit/seo-audit-{domain}-{YYYY-MM-DD}/
⚠️ The Website Audit/ parent already exists and holds every prior audit — write there, NOT in the repo root. This was violated on 2026-08-06 (audit-client.example) and cost a 149 MB directory move plus an artifact URL rescue. Create the full path before crawling, not after.
3. Show cost estimate to user (based on detected page count):
- 8 core agents (Opus): ~$8
- With conditional agents: ~$15 max
- Await user confirmation if over $10
Phase 2: CRAWL
python -m scripts.audit.crawl_site "$URL" --max-pages 500 --concurrent 5 --delay 1.0 --output ./"Website Audit"/seo-audit-{domain}-{date}/crawl-results.json
Configuration:
- Max pages: 500 (user-overridable via --max-pages)
- Respect robots.txt: Yes
- Follow redirects: Yes (max 3 hops)
- Timeout per page: 30 seconds
- Concurrent requests: 5
- Delay between requests: 1 second
Phase 3: DETECT BUSINESS TYPE
python -m scripts.audit.business_type_detector ./"Website Audit"/seo-audit-{domain}-{date}/homepage.html --url "$URL" --json > ./"Website Audit"/seo-audit-{domain}-{date}/business-type.json
Business types: SaaS | E-commerce | Local (brick-and-mortar/SAB/hybrid) | Publisher | Agency | Other
Phase 4: SPAWN SUBAGENTS (parallel)
Always spawn (8 agents):
| Agent | Responsibility | Key Scripts |
|---|---|---|
audit-technical | Crawlability, indexability, security, URL structure, JS rendering | crawl_site, fetch_page, parse_html |
audit-content | E-E-A-T scoring (80 items), readability, thin content, YMYL detection | core_eeat_scorer |
audit-schema | JSON-LD detection, validation, deprecated type veto (T09) | schema_validator |
audit-sitemap | XML sitemap structure, coverage gaps, quality gates | fetch_page |
audit-performance | LCP, INP, CLS via PageSpeed Insights + CrUX field data | pagespeed_check |
audit-visual | Desktop + mobile screenshots, above-fold analysis, responsive check | capture_screenshot, render_probe |
audit-geo | AI crawler access, llms.txt, citability (134-167w optimal), brand signals | fetch_page |
audit-sxo | SERP backwards analysis, page-type mismatch detection, persona scoring | fetch_page, parse_html |
Conditionally spawn (7 agents):
| Agent | Condition | Responsibility |
|---|---|---|
audit-local | business_type ∈ {local_*} | GBP signals, NAP consistency, reviews, local schema |
audit-maps | local detected AND geoapify/dataforseo key available | Geo-grid, competitor radius |
audit-google | Google API key configured | CrUX field data, GSC indexation, GA4 traffic |
audit-backlinks | Always (CC free); enriched if Moz/Bing keys | DA/PA, referring domains, toxic links |
audit-cluster | /blog detected OR pillar signals | Semantic clustering, cannibalization |
audit-drift | Previous baseline exists for this domain | 17-rule drift comparison |
audit-ecommerce | business_type == "ecommerce" | Product schema, marketplace intel |
Each agent writes results to: ./"Website Audit"/seo-audit-{domain}-{date}/modules/{name}.json
A negative rendering claim needs a time-series, and a real browser to settle ties
Never rate a "does not render / shows zero / is blank / is missing" finding above medium from a single screenshot. On 2026-08-06 two such findings shipped as critical on audit-client.example and both were wrong — a headless capture had caught a slow fadeInUp animation before it started, and that frame was read as the final state.
Measured on that page: headless, 10s wait, 6 runs → 5 never started; headed, 10s wait, 3 runs → all animated, 1 still mid-flight. So the failure mode is a single capture landing mid-animation, not "headless is broken" — sampling over time is the fix, and the real browser is the tie-breaker.
python -m scripts.audit.render_probe "$URL" --selector "[data-to-value]" --selector ".hero h1" --check
It reports three columns that must never be collapsed into one claim:
| Column | Audience | If content is missing here |
|---|---|---|
static (no JS) | text-extracting crawlers, AI fetchers, |