
frontend-craft
A design skill for Claude Code and OpenAI Codex, built on a graded corpus of 826 real sites.
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up frontend-craft (claude-skill project) into my current project. Found on https://claudeers.com/frontend-craft Repo: https://github.com/jaydendavisnc/frontend-craft Homepage/docs: — Detected install method: claude-skill → # copy this skill into .claude/skills/frontend-craft/ Category: other. Platforms: api, web, mobile. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: unknown; community-verified: false. Confirm the source before running anything.
# copy the skill dir into your project: # .claude/skills/frontend-craft/ (or ~/.claude/skills/frontend-craft/ for all projects)
git clone https://github.com/jaydendavisnc/frontend-craft
// compatibility
| Platforms | api, web, mobile |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | — |
| Pricing | open-source |
| Language | HTML |
Frontend Craft
A design skill that makes AI models build frontends like a professional team. It works in Claude Code and OpenAI Codex.
Left alone, models collapse onto a few favourite pages. This skill replaces their habits with a measured corpus: 826 real sites analysed and graded, broken into 8,917 tagged element examples and 1,570 element variants with exact specs.
Install
Claude Code:
git clone https://github.com/jaydendavisnc/frontend-craft ~/.claude/skills/frontend-craft
OpenAI Codex:
git clone https://github.com/jaydendavisnc/frontend-craft ~/.codex/skills/frontend-craft
Then ask for a frontend as you normally would. Python 3.8 or later runs the sampler. Node 22 or later and Chrome run the audit tools.
How it works
- Read the person.
references/intent.mdturns the brief into decisions: every named site, colour, font and section is kept, and mood words such as "clean" or "premium" are decoded rather than taken literally. What the person asked for outranks everything the skill would otherwise choose. - Get the pictures.
scripts/photos.pyfinds reusable photographs,scripts/cutout.swiftcuts out subjects andscripts/imgset.pygrades them as a set. - Deal.
scripts/direction.pydeals a different direction each run, filling only what the brief leaves open: a register, a type system, a palette, a voice and one move for each job the page has to do, taken from real sites that do not look like templates. - Have ideas.
references/taste.mdgives the model an idea engine and better directions for every part of a site, with first principles for scroll, typography and components. Every scene is planned as a card and rated on an ambition ladder before any code. - Build. Real copy first, then the first scene on its own, then the rest, then one motion score for the whole page.
- Audit.
scripts/shoot.mjsruns the page in a real browser,scripts/audit.mjsscores how template-shaped it is,scripts/lint.mjsflags generated copy, andscripts/slop.mjsnames every element to remove. The model fixes it and goes again.
It also covers work that is not a web page. references/beyond-web.md has first principles for slide decks, documents, dashboards, mobile apps, email, social graphics, posters, data graphics, identities, game interfaces, motion and packaging, and the sampler deals a taste hand for those instead of page sections.
GPT models also receive coaching, worked examples from top-scoring pages, specific business angles with bespoke components, and forced rebuild rounds.
Results
Scored blind by panels of Claude models, out of 10, with real top-tier sites mixed in. See RESULTS.md for the numbers and their limits. See CHANGELOG.md for what changed.
| Test | Without | With | Real sites |
|---|---|---|---|
| GPT-6, 6 briefs never tuned on | 5.67 | 6.72 | 7.92 |
| Claude Opus, 12 tuning briefs | 6.33 | 7.93 | 7.85 |
| GPT-6, 12 tuning briefs | 5.21 | 5.99 | 7.85 |
Version 1.5 against 1.0, Claude Opus, the same 12 briefs rebuilt and judged blind in a fresh panel: 7.17 for 1.0, 7.46 for 1.5 (gain +0.29, 95% interval +0.06 to +0.52, better on 10 of 12), with real sites at 6.69 in the same panel. Our first draft of 1.5 lost on all twelve; RESULTS.md has every round.
There is no clean held-out result for Claude.
// faq
What is frontend-craft?
A design skill for Claude Code and OpenAI Codex, built on a graded corpus of 826 real sites.. It is open-source on GitHub.
Is frontend-craft free to use?
frontend-craft is open-source, so it is free to use.
What category does frontend-craft belong to?
frontend-craft is listed under other in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/frontend-craft)
// retro hit counter
[](https://claudeers.com/frontend-craft)
// reviews
// guestbook
// related in Other
Anti-AI-slop design skill for Claude Code, Cursor, and Codex.
Huashu Design · HTML-native design skill for Claude Code · Claude Code 里 HTML 原生的设计 skill · 高保真原型 / 幻灯片 / 动画 + 20 设计哲学 + 5 维评审 + MP4 导出 · Agent-agnostic
Open source Ghostty-based macOS terminal with vertical tabs and notifications for AI coding agents. Built for multitasking, organization, and programmability.