design-apply

name	design-apply
description	Generate style-consistent frontend code using a saved design system (tokens.json + DESIGN.md) from the global library or a local DESIGN.md. Loads the design system + semantic roles + essence, re-screenshots the source for visual calibration, generates code constrained to the palette, typography, and component language, then runs a closed visual-verification loop (render → measure → iterate) instead of an eyeball checklist.
argument-hint	用 <风格名> 做一个 <页面描述>
metadata	{"requires":"a saved design system (run design-distill first) + a browser for the visual loop (degrades to token-only)","promptSignals":{"phrases":["用 ... 做","build ui","make a ... like","generate a page","match this brand","用 linear 做","做个 ... 页面","apply the design"]}}
retrieval	{"aliases":["generate UI matching a saved style","build a page in a distilled design system","apply a design system to new code"],"intents":["generate frontend code that matches a saved design system","build a page that looks like a specific brand/site","produce style-consistent UI from tokens.json + DESIGN.md"],"examples":["用 linear 做个博客主页","make a settings page like Stripe","用 vercel 做个登录页"]}

Design Apply

Step 1: Load Design System

Resolve the design system based on priority:

The library lives at ~/.config/design-distill/ (override with $DESIGN_DISTILL_HOME). Resolve with ordinary file ops — there is no CLI.

Priority 1: Explicit style name from user input

If the user mentions a style name (e.g., "用 linear 做个博客主页"), read its structured set + view:

DIR="${DESIGN_DISTILL_HOME:-$HOME/.config/design-distill}/<name>"
cat "$DIR/tokens.json"   # canonical: semantic roles + essence (preferred)
cat "$DIR/DESIGN.md"     # rendered view

If the directory doesn't exist, list available styles and report:

ls -1 "${DESIGN_DISTILL_HOME:-$HOME/.config/design-distill}"

Priority 2: Local ./DESIGN.md

If no explicit style name but ./DESIGN.md exists in the current project:

cat ./DESIGN.md

This is the project's own design system (e.g., the user is building a product and has documented their own design language).

Priority 3: No style available

If neither an explicit name nor a local DESIGN.md:

ls -1 "${DESIGN_DISTILL_HOME:-$HOME/.config/design-distill}"

Display available styles and ask the user to choose one. If none exist, suggest running design-distill first.

---

Step 2: Re-Screenshot Source for Visual Calibration

Extract the source_url field from the loaded DESIGN.md.

If source_url is present

Use the bundled screenshot primitive (self-contained — no /browse dependency) to capture the source:

node scripts/screenshot.mjs <source_url> --out <tmpdir> --viewports desktop --schemes light
# For a specific page type (e.g., pricing), re-run with that subpage URL

Load the captured images with the Read tool — look at the reference images before designing.

If node scripts/setup.mjs --probe reports tier: "token-only" (no browser), skip screenshots and design from the DESIGN.md values, clearly noting that visual calibration was skipped.

Why re-screenshot? Text descriptions drift toward stereotypes ("developer tool = dark mode"), while screenshots are facts. Vercel's default is light mode.

If re-screenshot fails (404, timeout, site unreachable)

Fallback to archived screenshots:

Check if archived screenshots exist:

ls ~/.config/design-distill/<name>/screenshots/ 2>/dev/null

If archived screenshots exist:

• Load them for visual calibration
• Warn: "Using archived screenshots (source site unavailable)"

If no archived screenshots:

• Generate based on document values only
• Warn: "Source site unavailable, no archived screenshots. Generating from document values only."

If no source_url

The design system was extracted from a local project. Generate based on document values only, without screenshots.

---

Step 3: Style Calibration

Prefer the structured tokens.json (next to DESIGN.md) when present — it carries the semantic roles (color-primary, color-surface, color-text, …) and the essence (the ≤5 defining traits). Generate against the semantic roles, not raw hex guesses.

Looking at the screenshots (if available), tokens.json, and DESIGN.md, identify:

• Essence first: the ≤5 traits from tokens.json.essence — these define the look; honor them above incidental detail
• Semantic roles: which color plays primary / surface / text / status — map UI elements to roles, not to arbitrary palette entries
• Core visual facts: background light or dark? primary font? CTA color?
• Anti-patterns — what this design would never do (from the document's Anti-patterns section)

---

Step 4: Generate

Produce frontend code, HTML, or design artifacts based on the user's request.

Constraints:

• Map elements to semantic roles (primary CTA → color-primary, page bg → color-surface, body → color-text) — colors come from the palette, no new colors
• Fonts must match the document's font families
• Reference the screenshot's actual proportions and visual density
• Component shapes follow the document's component vocabulary and radius-button/radius-card

---

Step 5: Closed Visual-Verification Loop

Do not rely on an eyeball checklist. Render the generated output, measure it against the saved design system, and iterate on the concrete deltas until it converges.

Render target: a static artifact via its file:// path; a framework project via a local dev server URL.

node scripts/verify.mjs ~/.config/design-distill/<name>/tokens.json <output.html|url> --out <evidence-dir>
# prints { score, pass, totalDeltas, instructions[], evidence }; exit 0 = pass

Loop (bounded + convergent — defaults: threshold 85, cap 3 iterations, stop on no-improvement):

1. Run verify.mjs. If pass (score ≥ threshold) → done; present the result + evidence.
2. Else apply the returned instructions (e.g. "color-primary should be #5e6ad2 (got #ff0000)") to the code and re-run.
3. Stop at the iteration cap or when the score stops improving; present the best round with its remaining deltas.

Degradation: if verify.mjs reports visualVerification: false (no browser), it ran a token-only check and skipped screenshots — tell the user visual verification was skipped rather than implying it passed visually.