Agent Skills

Website Frontend Visual Replication

Origin: Adapted from zai-org/GLM-skills/glmv-web-replication (Apache 2.0). Agent-browser replaced with Playwright MCP (built into OpenCode).

Prerequisites

This workflow uses Playwright MCP (built into OpenCode via playwright_browser_* tools). No external dependencies needed.

Authorization Gate (MUST execute first)

Before proceeding, MUST ask the user:

“Do you own this website, or do you have explicit written permission from the owner to replicate it? Unauthorized replication may violate copyright, terms of service, or applicable law.”

• If confirmed → proceed
• If cannot confirm → STOP. Suggest building an original design instead.

Scope & Limitations

Included	NOT Included
Page layout & visual styling	Backend / server-side logic
Navigation structure	Databases & data stores
Publicly visible text & images	Authentication systems
CSS/design tokens	API business logic
Client-side interaction patterns	Non-public content
Static asset files (images, fonts)	Credentials, secrets

Data handling rules:

1. Never scrape behind a login wall
2. Never collect credentials, API keys, or PII
3. Never reproduce copyrighted content verbatim unless user holds rights
4. Respect robots.txt and rate limits

Core Idea

1. Recursively explore every public page, record visual content, interactions, and assets into a structured “blueprint” with nested folders representing navigation relationships.
2. Build a frontend visual replica based on that blueprint.
3. Visually compare and revise.

Blueprint Structure

blueprint/
├── _meta.md                      # Site metadata
├── _sitemap.md                   # Sitemap
├── _assets/                      # Global assets (fonts, favicon)
├── _navigation_graph.md          # Navigation graph (Mermaid)
├── home/                         # Homepage
│   ├── _page.md                  # Page blueprint
│   ├── _full.png                 # Full-page screenshot
│   ├── _scroll_00.png ~ N.png    # Scroll sequence
│   ├── _interactions.md          # Interaction records
│   ├── _interactions/            # Interaction state screenshots
│   ├── _assets/                  # Page-specific assets
│   ├── products/                 # Child pages (reachable from here)
│   │   ├── _page.md
│   │   ├── _full.png
│   │   └── ...
│   └── about/
│       └── ...

Replication Workflow

Step 1: Initialize

mkdir -p blueprint/_assets

Open the target site with Playwright:

playwright_browser_navigate url="<target URL>"
playwright_browser_resize width=1920 height=1080
playwright_browser_wait_for time=3

Write blueprint/_meta.md:

# Website Replication Blueprint
- Target website: <URL>
- Exploration date: <date>
- Viewport size: 1920×1080

Step 2: Recursively collect pages

For every page, execute this procedure:

2.1: Capture and analyze

playwright_browser_take_screenshot fullPage=true filename="blueprint/<page>/_full.png"
playwright_browser_snapshot    ← get interactive elements

2.2: Download assets

Collect all image/video/font URLs from the page snapshot, download with curl:

curl -o blueprint/<page>/_assets/image.jpg "<image_url>"

2.3: Traverse interactions

For each interactive element found in the snapshot:

# Hover
playwright_browser_hover target="<element>"
playwright_browser_take_screenshot filename="blueprint/<page>/_interactions/hover.png"

# Click (may trigger navigation)
playwright_browser_click target="<element>"
playwright_browser_wait_for time=2
playwright_browser_take_screenshot filename="blueprint/<page>/_interactions/click.png"

# If navigation happened, capture the new page as a child
playwright_browser_snapshot   ← check if URL changed

# Go back
playwright_browser_navigate_back

2.4: Scroll and repeat

playwright_browser_press_key key="PageDown"
playwright_browser_wait_for time=1
playwright_browser_take_screenshot filename="blueprint/<page>/_scroll_N.png"

Continue until bottom of page.

2.5: Document the page

Write blueprint/<page>/_page.md:

# <Page Name>
- URL: <path>
- Source: <how we got here>

## Section Structure
| No. | Section Name | Layout Pattern | Content Type |
|------|--------------|----------------|--------------|
| 1 | ... | ... | ... |

## Screenshots
- Full: !_full.png
- Scrolls: !_scroll_00.png ... !_scroll_N.png

## Outbound Navigation
| Trigger | Method | Target | Child Folder |
|---------|--------|--------|--------------|
| ... | click | /path | ./child/ |

## Assets
| File | Purpose |
|------|---------|
| _assets/img.jpg | ... |

Write blueprint/<page>/_interactions.md:

# Interactions

| Element | Trigger | Behavior | Navigation | Screenshot |
|---------|---------|----------|------------|------------|
| ... | click | ... | Yes → ./child/ | ![](_interactions/click.png) |

Step 3: Generate summaries

After all pages collected:

blueprint/_sitemap.md

# Sitemap
home/                          # /
├── products/                  # /products
│   ├── product-detail/        # /products/:id
│   └── category/              # /products/category/:slug
├── about/                     # /about
└── blog/                      # /blog
    └── blog-post/             # /blog/:slug

blueprint/_navigation_graph.md

# Navigation Graph
graph LR
    Home["home /"] -->|Nav - Products| Products["products /products"]
    Home -->|Nav - About| About["about /about"]
    Products -->|Card click| Detail["product-detail /products/:id"]

Step 4: Frontend Visual Replication

Based on the blueprint, build a frontend replica:

• Match layout, colors, typography, spacing from screenshots
• Implement navigation between pages
• Use downloaded assets
• Match interaction patterns (hover states, filters, etc.)

Recommended stack: HTML + CSS + vanilla JS, or React/Tailwind if complex.

Step 5: Visual Comparison & Revision

1. Start the replica locally
2. Open both original and replica side by side (two browser tabs)
3. Compare page by page: layout, colors, typography, navigation
4. Fix discrepancies
5. Repeat until satisfied

Key Rules

1. Authorization first — never start without user confirmation
2. Public pages only — no login-protected areas
3. No credential handling — redact any found credentials
4. Frontend only — no backend replication
5. Folders = navigation — if page A links to page B, B is a subfolder of A
6. Every folder needs: _page.md, _full.png, _interactions.md, _interactions/, _assets/
7. Screenshots from real site — never describe from memory
8. Interactions genuinely triggered — hover, click, focus, one by one
9. Download assets — curl for images, Playwright eval for SVGs
10. Record only — document results without quality judgments
11. Keep files updated — real-time updates after each page exploration
12. Visible elements — scroll into view before interacting
13. One action per Playwright call — no combining multiple commands
14. External links — record URL + trigger element, don’t deep-explore
15. Both full and scroll screenshots — overall + detail reference

Boundaries

• Do not use to copy proprietary content or bypass access controls; use it only for authorized references.

Evidence and completion gate

Ground recommendations in the supplied repository or brief. State assumptions, missing inputs, and unresolved risks. Return a concrete artifact or checklist with an owner/action for each open item, then verify that the result answers the requested goal rather than merely repeating the framework.

This site is open source. View this page on GitHub.