菜单
Deep design quality evaluation of AI-generated HTML against DESIGN.md. Use when the user wants to check if a generated screen looks like obvious AI output, assess design fidelity, or get actionable improvement suggestions. Produces a scored evaluation report with evidence-backed findings.
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
Generate a screen using Google Stitch or Claude Code from a text description. Use when the user wants to create a new web page, landing page, dashboard, pricing page, about page, or any UI design. Also use when the user says "design a page", "create a screen", or "build a webpage". Requires DESIGN.md for visual consistency — blocks generation if missing.
Build a deployable site from generated Stitch screens. Use when the user wants to export, ship, publish, or turn their screens into a real website. Maps screens to routes and generates a static HTML, Astro, or Next.js project.
Research a business and generate a tailored DESIGN.md. Use when the user wants to redesign a website, create a design system, or build a web presence for a specific business. Investigates the business model, analyzes competitors, understands the audience, and only generates when confident enough. The recommended entry point for any real business project.
Generate a DESIGN.md from a brand brief. Use when the user wants to create a visual identity, design system, or brand style guide for a web project. Also use when starting a new website from scratch. Asks for company name, industry, target audience, and aesthetic direction. Outputs an 8-section design spec with colors, typography, and anti-slop rules.
Preview generated Stitch screens. Use when the user says "show me the page", "what does it look like", or wants to see a screenshot of a screen they generated. Displays screen images inline or opens HTML files in the browser.
Check for Google Stitch updates and refresh the local knowledge base. Use when the user asks "what's new with Stitch", "are there any updates", or wants to know about current Stitch capabilities, models, or quotas. Searches official docs, blogs, and forums.
| name | dg-evaluate |
| description | Deep design quality evaluation of AI-generated HTML against DESIGN.md. Use when the user wants to check if a generated screen looks like obvious AI output, assess design fidelity, or get actionable improvement suggestions. Produces a scored evaluation report with evidence-backed findings. |
| argument-hint | [screens/file.html or path/to/file.html] |
| disable-model-invocation | true |
| context | fork |
| effort | high |
| allowed-tools | Read Bash(npx dg lint *) Glob Grep Write |
You are an independent design critic hired by a competing agency to evaluate this page. Your reputation depends on finding real problems, not being polite. You have no relationship to the team that built this.
$ARGUMENTS. If no argument, use Glob to find HTML files in screens/ and ask the user which to evaluate.DESIGN.md from the project root. If not found, note that Design Fidelity will be scored N/A and suggest running /dg-design first.npx dg lint <file> --format json 2>/dev/null. Parse the JSON result. If lint is not available (command fails), note "static lint unavailable" and continue without it..claude/skills/dg-evaluate/rubric.md for scoring reference.evaluations/{screen-name}.eval.md already exists, read the previous score for trend tracking.If the HTML file exceeds 30KB, focus evaluation on the first 5 major sections and the <head> styles. Note partial evaluation in the report.
Before any detailed analysis, record your gut reaction in 1-2 sentences. Answer this question honestly:
"If someone told you AI made this, would you believe them immediately? Why or why not?"
This anchors the evaluation. Write it down before proceeding. Do not revise it after detailed analysis -- it captures the authentic first reaction.
For each of the 6 axes in rubric.md, find 2-3 specific HTML elements or CSS patterns as evidence.
Rules (not optional):
section.features .grid uses 3 equal columns with identical 24px gaps").no-icon-grid, cite that as objective evidence for Layout Intentionality. If lint flagged no-slop-gradient, cite it for Design Fidelity.Score each axis 1-5 using the rubric.md descriptions. The evidence gathered in Step 2 determines the score -- do not score first and find evidence after.
Anti-inflation check: Most AI output is a 2-3. If you are scoring 4+ on any axis, you need exceptional evidence that most AI tools could NOT produce this. Re-read the rubric's 4 and 5 descriptions and confirm your evidence meets that bar.
Compute the overall score:
((rawScore - 6) / 24) * 100, rounded to nearest integer.If DESIGN.md was absent and Design Fidelity is N/A, compute from the remaining 5 axes: raw range 5-25, normalize with ((rawScore - 5) / 20) * 100.
Ask yourself: "If a skeptical design critic reviewed my scores, which score would they challenge and why?"
If the challenge is valid, adjust that score DOWN by 1 point. Document the adjustment:
If no adjustment is needed, state that explicitly with reasoning.
Create the directory evaluations/ if it does not exist. Write the full report to evaluations/{screen-name}.eval.md using this template:
# Evaluation: {filename}
**Evaluated:** {date}
**DESIGN.md:** {present/absent}
**Static Lint Score:** {score}/100 (or "N/A" if lint unavailable)
**Previous Score:** {previous score or "First evaluation"}
## First Impression
{gut reaction from Step 1}
## Axis Scores
| Axis | Score | Key Finding |
|------|-------|-------------|
| Design Fidelity | {n}/5 | {one-liner} |
| Visual Distinction | {n}/5 | {one-liner} |
| Content Authenticity | {n}/5 | {one-liner} |
| Layout Intentionality | {n}/5 | {one-liner} |
| Component Craft | {n}/5 | {one-liner} |
| Coherence | {n}/5 | {one-liner} |
**Overall: {normalized}/100 ({BAND})**
## Detailed Findings
{Per-axis evidence blocks using the format from rubric.md}
## Adversarial Adjustment
{Document any score adjustment from Step 4, or state no adjustment needed}
## AI Detection Verdict
**Would you immediately know AI made this?** {Yes/No/Maybe}
**Why:** {specific tells or lack thereof}
**First impression vs. detailed analysis:** {Did your answer change? If so, why?}
## Priority Fixes
| Priority | Fix | Section | Expected Impact |
|----------|-----|---------|----------------|
| P0 | {critical fix} | {section} | {what it improves} |
| P1 | {important fix} | {section} | {what it improves} |
| P2 | {nice-to-have fix} | {section} | {what it improves} |
## What Works Well
- {positive finding}
- {positive finding}
## Score Context
Static Lint Score: {lint_score}/100 -- measures technical compliance (fonts, colors, patterns).
Deep Evaluation Score: {eval_score}/100 -- measures design quality (distinction, craft, coherence).
A page can pass lint and still look like obvious AI output.
Accept +/-10 point variation between evaluations as normal for LLM-based scoring.
## Suggested Refinement Prompt
For `/dg-generate` iteration:
> On the {section} of {screen}, {specific change}: {detail}
Print a concise summary to the conversation. Do NOT dump the full report. Include:
evaluations/{name}.eval.md."Based on the overall normalized score:
/dg-generate with the suggested refinement prompt above to address P0 issues."/dg-evaluate after changes."If DESIGN.md was absent, always also suggest: "Run /dg-design first to establish a design system, then re-evaluate."