notque/vexjoy-agent / skills/frontend/jev-design
jev-design
VerifiedDesign React interfaces with the full shadcn/ui catalog. The agent writes a brief, Jev picks components and a style recipe from bounded catalogs, then the agent implements, renders, and reviews against concrete checks.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 7.4 KB |
| references/brief-writing.md | markdown | 3.3 KB |
| references/catalog-v1.json | config | 2.3 KB |
| references/catalog-v2.json | config | 48.9 KB |
| references/decision-card.md | markdown | 2.5 KB |
| references/jev-images.md | markdown | 3.9 KB |
| references/visual-review.md | markdown | 3.3 KB |
| scripts/plan.py | script | 28.2 KB |
| tests/test_plan.py | script | 17.3 KB |
references/jev-images.md
View on GitHub# Review renders with Jev images
Send screenshots to Jev for phase 4 review instead of describing the page in text. Jev reads the picture and your question together. Shared API, limits, token cost, and routing: [images.md](../../../meta/building-with-jev/references/images.md). Image judgments steer review rounds. They never replace the `visual-review.md` checks or a measured number, and they are never the only gate.
## Capture
TypeSafe's advice: the model does better with high-resolution images, so aim for at least 1,080 px on the long side, with JPEGs of about 512 KB to 2 MB as the sweet spot. Accept the token cost: a sharp image beats a cheap blurry one. Provide structured context beside it (measured numbers, element facts), not the picture alone.
| Image | Size | Why |
|---|---|---|
| Desktop, light | 1280 x 800 viewport, `deviceScaleFactor` 1, scrolled to the area under review | Long side 1280. About 2,700 tokens. |
| Mobile, light | 375 x 700 viewport, `deviceScaleFactor` 2 (750 x 1400 image) | Long side 1400, text stays sharp. About 2,700 tokens. A 375 x 812 shot at scale 1 is too small. |
| Dark variant | Same sizes, only when a dark finding is open | Skip otherwise. |
| Reference | A rated screen captured the same way | Anchors "good" to something real. |
- Capture viewports or section crops. Keep every side 280-1568 px; a full-page capture taller than 1,568 px is shrunk until text is unreadable. Do not shrink a capture to save tokens.
- Save as JPEG (quality about 85) when a PNG passes 2 MB.
- Scroll to the bottom first so scroll-reveal content has rendered.
- `JevImage` fixes small deviations.
## Where each image goes
| Image | Place | Reason |
|---|---|---|
| Desktop and mobile candidate | `state`, once | All questions in the request share it, so it is billed once. |
| Reference screen | `state`, with the owner's rating in text beside it | Every trait question compares against it. |
| Example of a failure (clipped button, low contrast) | The `criteria` of the Noul it defines (`true.examples`) | Only that question pays for it. |
## Question set
One Noul per trait, plus a separate "anything worse" Noul. The measured numbers (control heights, contrast ratios, horizontal scroll) go in state as text next to the images.
```python
from jev_images import JevImage
import jev_transport
state = {
"desktop": JevImage("desktop.png"), # candidate, 1280 px wide
"mobile": JevImage("mobile.png"), # candidate, 375 px at scale 2 (750 x 1400)
"reference": JevImage("ref.png"), # owner-rated screen
"reference_rating": "owner rates the reference 9 of 10",
"measured": {"primary_buttons_in_footer": 1, "horizontal_scroll_375": False},
}
questions = {
"primary_clear": {"type": "noul", "instructions": "Is the single primary action in `desktop` obvious within five seconds?"},
"mobile_primary_visible": {"type": "noul", "instructions": "Is the primary action in `mobile` visible without scrolling?"},
"hierarchy": {"type": "noul", "instructions": "Does `desktop` show a clear heading, section and body hierarchy like `reference`?"},
"worse": {"type": "noul", "instructions": "Does `mobile` show clipped, overlapping, or unreadable elements?"},
"gap": {"type": "score", "instructions": "How far is `desktop` from `reference`?", "criteria": LADDER},
}
answers = jev_transport.evaluate(state, questions, timeout=90)["answers"]
```
- `LADDER` is a Score ladder with `summary` and observable `signals` (sizes, alignment, what is visible).
- Budget: two candidate images plus a reference come to about 8,100 tokens. Read `usage.input_tokens` for the real number.
- Image judgments are one signal among several; ask one trait per question and keep the measured numbers beside the picture.
- Report each answer as a finding with a location and width, as `visual-review.md` requires. A Noul in 0.4-0.6 is not an answer: crop to the area in doubt and ask again.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |