Visual Design Pipeline
A back-half product-design pipeline: given "user scenario + what to display" as input, drive it to a high-fidelity chosen version.
Basic info
- Name
- Visual Design Pipeline
- Description
- 当用户要设计一个页面 (UI/视觉/产品层, 不是只写代码) 或修改已有视觉设计时, 驱动三段流水线 (显示信息 → 显示框架 → 风格+实践), 每页独立, 先文档后 HTML. 输入 = "已知用户场景 + 已知要显示什么", 输出 = 高保真选定版. 不用于: 文案小改 / 单 bug 代码修复 / 纯后端.
visual-pipeline
A back-half product-design pipeline: given "user scenario + what to display" as input, drive it to a high-fidelity chosen version.
The unit is a page. One page runs the pipeline once. Three stages in strict order: display info → skeleton → style + practice. Every stage writes docs before HTML, no stage-skipping, and every change goes through a routing decision.
Code implementation and everything after it are out of scope.
When to use it
- The user says "design / build / change page X" and the ask is on the visual / product layer
- Input has already converged to "user scenario + what to display", now waiting to be turned into pixels
- One page one run, multiple pages multiple runs, each page judges its own style
Not for: pure copy tweaks / single-bug code fixes / pure backend work / requirement-discovery stage where the ask is still unclear.
Comparison with similar skills
Drafted by tranfu-publish, signed off by the author. Helps a reader decide horizontally which one to install, or jump to a better-fit sibling.
Inside the company library
- ui-ux-pro-max — a UI/UX material library (67 styles / 96 palettes / 57 font pairings); this skill differs: it's a material library called inside the "style stage"; this skill is process discipline, governing stage order and change routing
- claude-design-system — a mirror of Anthropic's internal design-tool system prompt (a reference for HTML artifacts); this skill differs: it's a prompt reference, whereas this skill is an executable pipeline that enforces non-skippable stages
Outside world
- claude-wireframe-skill — one-shot output of five B&W wireframes plus parallel color variants; this skill differs: that one is "many parallel options", this one is "single page in depth with reversible changes"
- claude-design-skill — a replica of the Claude.ai internal prompt, produces HTML artifacts (decks / landing / animation); this skill differs: that one covers a wide range (including slides and posters), this one is scoped to page design with a mandatory three-stage order
Unique value of this skill
- Mandatory stages, docs before HTML, so visual decisions leave a traceable paper trail
- Any change is routed to its owning stage first, preventing the historical blocker from coming back
- Each page has its own style stage, protecting per-page choices from being flattened by a project-wide tone
Tips for using it
Drafted by the author. Helps a reader get started vertically.
Material lessons
- We tried "one-shot pile of HTML, let the user pick" — decisions had no docs, no way to roll back, and round two fell apart
- We tried "one style fits all pages" — special per-page scenarios got flattened, and after a few user rejections there was no anchor left
Recommended usage
- First confirm the page output directory with the caller (e.g.
web-goal-docs/<page>/) - Run all three stages on the first page to set a baseline; later pages can reference "inherited from the project shared design language"
- When the user says "wrong", route first, then edit — don't jump straight into the HTML; this is the core discipline
Known limits
- Doesn't make requirement-layer decisions (user scenario, field priority) — inputs must be locked upstream
- Doesn't emit code;
03-selected.htmlis the ground truth, decisions made at code layer can't roll back into the style stage - Everything is in Chinese, artifacts forbid short labels like S1/P0/R1 (may not match the conventions of an English-first project)
Version info
Local public Skill catalog snapshot, showing only public-safe fields.
Skill files
(20)SKILL.md
SKILL.md · Markdown