A structured 100-point visual audit system for B2B websites, built as a Claude Skill.
Turn subjective client feedback into objective scores. Evaluate any website across 5 conversion-focused dimensions and get actionable findings in minutes — not hours.
If you build websites for B2B clients — especially in manufacturing, export, or tech — you know this situation:
"I don't like the colors."
"Can we make it look more like our competitor?"
"It just doesn't feel right."
These subjective requests lead to endless revision cycles. This skill gives you data to anchor the conversation: a 100-point score, a 20-item rubric, and industry benchmarks — so changes are evaluated on impact, not opinion.
Given any URL, the skill:
- Fetches the page content via
web_fetch(with fallback toweb_search) - Extracts 40+ structural, trust, content, and technical signals
- Scores across 5 weighted dimensions (20 sub-items, each 1–10)
- Classifies findings into Block / Improve / Keep tiers
- Outputs a structured report with dimension bars, per-item reasoning, and prioritized action list
| Dimension | Weight | What It Measures |
|---|---|---|
| Visual Hierarchy | 25% | Can visitors find the value prop in 3 seconds? |
| Brand Consistency | 20% | Is the visual identity coherent and industry-appropriate? |
| Conversion Guidance | 25% | Do CTAs, trust signals, and forms drive inquiry? |
| Readability | 20% | Is content scannable for a busy B2B buyer? |
| Performance Perception | 10% | Does the page feel fast and uncluttered? |
| Score | Verdict |
|---|---|
| 85–100 | ✅ 优秀 — Ready to ship |
| 70–84 | 🟡 良好 — Minor fixes, then ship |
| 55–69 | 🟠 待改进 — Targeted improvements needed |
| < 55 | 🔴 重构级 — Significant rework required |
- Download
website-visual-scorer.skillfrom this repo - In Claude.ai, go to Settings → Skills → Install Skill
- Upload the
.skillfile - Start a new conversation and say:
Evaluate https://yoursite.com
Claude will fetch the page, run the full 20-item audit, and return a structured report.
If you're building on the Claude API or Claude Code, copy the contents of website-visual-scorer/SKILL.md into your system prompt or skill directory.
See examples/ for real evaluation reports:
| Site | Industry | Score | Report |
|---|---|---|---|
| GNSource | Defense / Precision Tech | 88 | view |
| TTI Fiber | Fiber Optic Manufacturing | 79 | view |
| BisonConvey v2 | Industrial Conveyor Mfg | 76 | view |
| RapidDirect | CNC Manufacturing | 68 | view |
| BisonConvey v1 | Industrial Conveyor Mfg | 56 | view |
Provide 2+ URLs to get a side-by-side leaderboard:
Compare https://site-a.com and https://site-b.com
Output includes dimension-by-dimension delta and cross-site best practices.
website-visual-scorer/
├── SKILL.md # Main skill — execution protocol
└── references/
├── scoring-rubric.md # 20-item scoring criteria with band descriptions
└── industry-benchmarks.md # Benchmark template (add your own data)
| File | Status | Notes |
|---|---|---|
SKILL.md |
✅ Open | Full execution protocol |
references/scoring-rubric.md |
✅ Open | Complete 20-item rubric |
references/industry-benchmarks.md |
✅ Template | Empty template — populate with your own evaluations |
The scoring framework is fully open. The benchmark database (real scores from real sites) is not included — you build your own as you evaluate more sites. This is intentional: your benchmark data is your moat.
These findings are automatically flagged as blocking issues regardless of overall score:
Lorem ipsumplaceholder text anywhere on page- AI-generated image filenames exposed (
Gemini_Generated_Image_*,DALL-E-*, etc.) - Certification image belonging to a different company than the site owner
- Spelling errors in navigation menu items
- Primary CTA button pointing to
#(dead link) - Non-target-language alt text revealing incomplete internationalization
Using Claude Sonnet 4.6 (recommended model):
| Configuration | Cost per Evaluation |
|---|---|
| Standard | ~$0.07 |
| With Prompt Caching | ~$0.05 |
| Batch API | ~$0.03 |
| Haiku 4.5 (budget) | ~$0.02 |
At $0.05/eval with prompt caching: 1,000 evaluations/month costs ~$50.
Contributions welcome. See CONTRIBUTING.md for guidelines.
High-value contributions:
- New industry scoring adjustments (e-commerce, SaaS, healthcare, logistics)
- Additional Red Flag detection patterns
- Non-English language evaluation support (Japanese, Arabic, Spanish)
- Evaluation examples from more industries
Do not submit:
- Proprietary client data or real business benchmark numbers
- Evaluations of sites without owner permission
This skill was created by a team building AI-powered websites for B2B export companies. We use it internally to QA sites before delivery and to redirect subjective client revision requests toward data.
If you're doing similar work — building or auditing B2B websites — this framework was built for you.
MIT License. Use freely, commercially or otherwise. Attribution appreciated but not required.
See LICENSE for details.