菜单
Run a final review of accessibility, AI-template tropes, hierarchy/rhythm, and interaction states, then fix what's found. Use only when the user explicitly asks to polish the build, run final checks, or make it ready to ship.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Audit an HTML/UI Studio build for accessibility — contrast, semantic HTML, keyboard/focus, motion, forms — and fix what's found. Use only when the user explicitly requests an accessibility review or a polish-pass.
Scan a Studio HTML/UI build for generic AI-template visual tropes and fix them. Use only when the user explicitly asks whether it looks AI-generated, asks to remove the slop, or requests a polish-pass.
Create generative / algorithmic art — flow fields, particle systems, noise-driven compositions — rendered to an image with code. Use when the user wants art made with code, generative art, or parametric visuals.
Design a software/system architecture diagram (C4 context/container, or a sequence diagram) from the conversation and workspace material, delivered as a Mermaid `.mmd` diagram that renders natively in the panel. Trigger with "draw the architecture", "create a C4 diagram", "diagram the system", "show a sequence diagram", or whenever the user wants a technical system visualized.
Build a self-contained interactive HTML dashboard with ECharts, filters, KPIs, and tables. Use for executive overviews, operational monitoring, or analytical workspaces where several coordinated views need to respond to the same data.
Create beautiful static visual art — posters, design pieces, cover art — as .png or .pdf, driven by a design philosophy. Use when the user asks for a poster, a piece of art, a design, or any static, design-forward visual. Create original work, never copying existing artists.
| name | polish-pass |
| description | Run a final review of accessibility, AI-template tropes, hierarchy/rhythm, and interaction states, then fix what's found. Use only when the user explicitly asks to polish the build, run final checks, or make it ready to ship. |
An optional umbrella review for an interactive/visual build. A polished and an unpolished build are the same idea at different levels of care — the gap is what people actually notice. Don't run this on work that's still mid-structure (broken layout, missing sections, content the user is still iterating on) — say so and ask whether to polish now or after the structure settles.
When the agent tool is available and the build is large enough to justify delegation, launch the
four checks in parallel. Each briefing must be self-contained: pass the file path and tell the agent
to read it before checking. Ask for concrete findings with severity and evidence, not speculative
low-confidence lists.
read_skill accessibility-pass and audit contrast, semantic HTML, keyboard/focus, and
motion/forms.read_skill ai-slop-check and scan for gradients, emoji, trope cards, weak illustration,
overused fonts, the cream/serif/terracotta house style.read_skill hierarchy-rhythm-review and check size/color/weight/position/density hierarchy plus
spacing/type-scale rhythm.read_skill interaction-states-pass and inventory every interactive element for
hover/active/disabled/focus/loading plus transitions and action feedback.If agent is unavailable or the build is small, run the checks yourself. Delegation is an
optimization, not part of the quality bar.
Merge the four reports, de-duplicating overlaps (a removed focus ring shows up in both accessibility and interaction-states — one entry). Group into:
Fix directly. For a finding that's a false positive or genuinely out of scope (a third-party embed you can't touch), note it and move on rather than arguing with it.
After fixing, glance at the high-risk spots: did a contrast fix wash out the brand color, does a new focus ring overlap neighboring content, does the CTA actually feel primary now? Fix anything off.
End with one short summary: ready to ship / ready once the user reviews flagged calls / needs more iteration before polish is worth doing — plus blockers fixed, polish applied, and open decisions. No recap of what the user just watched you do.