Evidence-first website copy-remix skill for agent workflows
Catalog skill by wuxie888: Evidence-first website copy-remix skill for agent workflows Install into Cursor, Claude, Codex, or VS Code from this page or review the source for agent-skill-wannengcopy.
Ads
git clone https://github.com/wuxie888/agent-skill-wannengcopy
Install guides: Cursor skills, Claude Code skills, Codex skills, or VS Code Copilot skills.
Use this skill to turn a reference website into a usable site for the user's own product.
This skill supports lawful exact cloning when the user has rights, but defaults to experience remix: preserve the reference site's structure, interaction feel, motion language, and visual rhythm while replacing brand assets, copy, product logic, and media with original equivalents.
The main workflow is product-first. Use evidence recon from references/evidence-recon.md as the inspection and verification subsystem; do not let source mirroring or technical teardown override the goal of building the user's own product unless Mode 1 is clearly authorized.
Choose a mode before coding. Read references/copy-modes.md when mode or rights are unclear.
Pair the mode with a complexity level from references/evidence-recon.md when the reference is a live site:
Mode 2 + L4 = experience remix of an animation-heavy brand site
Mode 3 + L3 = product rebuild from a React/Vue/Next content site
Mode 4 + L5 = motion reconstruction of a WebGL/Canvas/Three.js effect
Mode 1 + L1 = authorized exact clone of a static site
Infer from local files when possible. Ask briefly only if the missing detail blocks progress.
reference website or screenshot:
target product:
one-line product promise:
target user:
core user action:
core workflow:
desired mode: authorized exact clone / experience remix / product rebuild / motion reconstruction / QA
evidence available: live URL / screenshots / repo / source files / none
must-use assets:
must-not-include:
Use the browser or available screenshots to inspect:
List transferable experience patterns and non-transferable surfaces.
For live websites, read references/evidence-recon.md and run the lightest honest recon path before implementation when feasible. Prefer real evidence over visual guesses:
Record conclusions as SOURCE, PARTIAL, or GUESS for complex motion claims.
Use this rule:
User owns or has explicit authorization -> Mode 1 can be used.
No authorization or unclear rights -> Mode 2/3 only.
Reference value is mostly motion -> also apply Mode 4.
Existing output has weird product mismatches -> Mode 5.
Do not carry over third-party logos, mascots, exact copy, screenshots, videos, or distinctive artwork unless Mode 1 is clearly justified.
If source code or deployed assets are available, still choose the legal/product mode before using them. "Publicly accessible" does not mean "safe to ship copied assets."
Create a compact table before implementation:
Reference module | Evidence | Experience to keep | Replacement for target product | Risk
Examples:
Hero cinematic video | screenshot + motion probe | large spatial impact and slow movement | original product engine visual | avoid reference video asset
SDK code panel | DOM + screenshot | technical theater and confidence | workflow status/log panel | avoid irrelevant developer APIs
Install cards | screenshot | three-column decision surface | user entry choices | avoid dependency install flow for consumers
Phone mockup | screenshot | secondary context/device contrast | result preview only if mobile matters | avoid decorative leftovers
If a reference module does not serve the target product, replace or remove it.
Read references/motion-levels.md whenever the reference includes advanced visual motion.
Use:
webm/mp4 plus native text overlay.Do not silently flatten a motion-heavy reference into a static background. If tooling prevents C/D, implement B and state the upgrade path.
For Level D or WebGL/Canvas-heavy references, use the evidence discipline in references/evidence-recon.md: find real source/source maps first, mark claims SOURCE/PARTIAL/GUESS, and avoid tuning guessed constants to hide unknown timing, shader, or state behavior.
For the target product, design around its real metaphor:
input -> processing/transformation -> output/result
Keep text, buttons, nav, and UI labels code-native whenever possible so the site stays editable and responsive. Use generated images or video only for visual scenes, not critical copy.
Follow the current repo's stack and patterns. Keep edits scoped.
Requirements:
For product remixes:
If recon produced reference artifacts, keep them under a non-shipped RECON/ or notes directory. Do not import scraped reference assets into production output unless Mode 1 or an explicit license allows it.
Use the in-app browser or Playwright-style browser checks when available.
Verify:
For motion, capture evidence beyond one screenshot: computed transform/opacity changes, video state, canvas pixel changes, or two screenshots separated by time.
Read references/verification-audit.md before final handoff when the task involved a live reference, exact clone, motion reconstruction, or a production-ready deliverable.
Before final, read references/product-logic-qa.md and check every visible section. Fix mismatches before reporting completion.
Report only high-signal details: