The FableGuide design guide

How to use Claude's design skills

A plain-language, end-to-end guide to designing genuinely distinctive pages with Claude — every design skill and command, how each model designs differently, and advanced ideas made simple. With six fully worked examples you can open, read, and steal from.

First, the honest part. "Claude design" is not a menu of named themes you pick from. The core skill, frontend-design, teaches a method for making deliberate, non-templated choices — it even names the looks AI defaults to and tells you to avoid them. The gallery below is six design directions we built with that method, labeled so you can learn from each — not official presets.
The toolkit

Every design skill & command — and how to invoke it

A skill is just packaged expertise Claude loads on demand. You invoke one by typing its slash-command, or simply by describing a task that matches it — Claude reaches for the right one.

1 · Making the design

frontend-design

/frontend-design  · or just ask for a page

The core skill. Everything in this guide grows out of it.

Turns Claude into "the design lead at a small studio" whose whole job is a look that could not be mistaken for anyone else's. It works in passes: brainstorm a token system (palette, type, layout, one signature element), critique that plan against the brief to strip out anything generic, then build. Invoke it before building or reshaping any page.

artifact-design

loads automatically before an Artifact

Design fundamentals tuned for shareable, single-file Artifacts.

When you ask Claude to "make me a page / dashboard / tool" and it publishes an Artifact (a private web page hosted on claude.ai you can share), this skill calibrates how much design effort the request deserves and keeps the result self-contained, responsive, and theme-aware in light and dark.

dataviz

/dataviz  · or ask for a chart

The one to reach for the moment a chart, dashboard, or stat tile appears.

A design system for data: a form heuristic (which chart for which question), a color formula with a runnable validator so series stay distinct and accessible, and specs for legends, axes, tooltips, sparklines and KPI tiles — coherent in both light and dark. Read it before writing the first line of chart code, in any library.

2 · Syncing to claude.ai/design (a real design system)

Beyond one-off pages, Claude can keep a shared design-system project on claude.ai/design — a living library of your components that your team and future sessions design against. Four commands run that loop:

/design

/design

The permission switch. Grants or revokes a Claude agent's access to your Design projects on claude.ai. Nothing reads or writes your design system until you've allowed it here.

/design-login

/design-login

Authorizes design-system access for a session that isn't already signed in through your claude.ai login — the one-time handshake that lets /design-sync talk to your project.

/design-sync

/design-sync

Pushes a React design system up to claude.ai/design. It runs a converter that renders your components to preview cards, then syncs them into the project incrementally — one component at a time, never a wholesale replace — so the library grows without clobbering. It always shows you the exact write/delete plan before touching anything.

/desktop

/desktop

Hands the current session off to Claude Desktop — handy when you want to keep designing in the desktop app's fuller visual surface without losing your place.

3 · Seeing & critiquing what you built

Design is a loop: build → look → fix. A picture is worth a thousand tokens, so give Claude eyes.

/layout

/layout <domain> [/extra-path …]

Screenshots a live site at desktop (1366×900) and mobile (390×844), plus an above-the-fold hero, then actually reads the images back and reports concrete problems — overlaps, cut-off text, broken tiles, weak contrast, anything that looks generic — and proposes fixes. Your fastest "does it really render?" check.

claude-in-chrome

describe a browser task

Drives your real Chrome — navigate, click, fill forms, read the console, and capture screenshots or even a GIF of a flow. Use it to critique a page in situ or record a walkthrough. (The /layout route above is the quick, headless version.)

Figma (MCP)

/figma-use  · or share a figma.com link

A two-way bridge between code and design. Pull a Figma frame into code, push a page you built into Figma, or build/extend a design system and its tokens — all from the same conversation. Trips automatically whenever you mention Figma or paste a figma.com URL.

/verify · /code-review · /simplify · /run

/verify · /code-review · /run

Not design skills — but the craft floor around a design.

/run launches the app so you can see a change working; /verify drives the changed flow end-to-end; /code-review catches the CSS-specificity traps that quietly cancel your styles; /simplify trims the cruft. A beautiful page that doesn't render isn't a design — it's a screenshot.

💡The publish path. To turn any of this into a shareable link fast, ask Claude for an Artifact — it writes the page, loads artifact-design, and returns a private claude.ai URL you can choose to share. For a page that lives on your own domain (like this one), Claude writes the file straight into your site.
Match the model to the job

How the different Claude models design

The skill is the same across models; the taste, ambition, and stamina differ. Bigger models hold a complex vision together across a long page and take braver, better-justified risks. Faster models are superb for a focused restyle but lean toward safe defaults if you don't steer them. Pick the model to fit the design job — and your usage budget.

Fable 5
claude-fable-5
Flagship · 2× cost
The most ambitious designer — richest signature moments, boldest justified risks, best at sustaining an elaborate maximal vision from hero to footer. Reach for it on a flagship page or a hard aesthetic brief. It costs ~2× Opus per token, so save it for the pieces that earn it.
Opus 4.8
claude-opus-4-8
The design workhorse
The everyday default. Strong taste, reliable restraint, coherent from plan to build — the right call for almost every real page. If you only learn one model's design habits, learn this one.
Sonnet 5
claude-sonnet-5
Fast & capable
Quick and genuinely good, especially for applying an established design system or restyling a single file. Give it a clear token system to work from; left fully open-ended it reaches for defaults sooner than Opus.
Haiku 4.5
claude-haiku-4-5
Mechanical edits
The scalpel. Perfect for token swaps, copy tweaks, spacing fixes, one small component — cheap and fast. Not the model to hand an open-ended "make it beautiful."

Rule of thumb: design the direction with Opus (or Fable for the flagship), then let Sonnet/Haiku apply and maintain it. Whichever model, invoke frontend-design and give it a real subject — the method matters more than the horsepower.

Advanced, made simple

The ideas that separate distinctive from templated

These are the concepts designers reach for. Each one is stated plainly, then in a sentence you can act on.

Design tokens

Decide each color and size once; name it; reuse the name.

Instead of scattering #1f7a6b through the file, define --accent once and refer to it everywhere. Change the brand in one line; keep the whole page consistent. Every example in the gallery is built from a 5-color, 2-typeface token set — that's why each feels like one coherent world.

A type scale with optical sizing

Pick a few deliberate sizes, and let display type be drawn for big.

Don't freehand font sizes. Choose a small ladder (body → h3 → h2 → h1) and jump between rungs. "Optical sizing" (the opsz axis in fonts like Fraunces) means the letters are redrawn to look right large vs. small — the storybook page pushes opsz to its max for warmth; a data table would pull it down for clarity.

Palette as named hex, not vibes

Write your colors down as 4–6 specific values before you build.

"Warm and friendly" is a wish; #f4d9a6 / #f2b134 / #3f8f6b / #e0623b / #c23e5a is a decision. Spend most of the palette on quiet, supporting tones and reserve one for emphasis — then use that one sparingly.

The signature element

Pick the one thing the page will be remembered by — and keep everything else calm.

The ribbon on Storybook, the constellation on Midnight, the misregistered overprint on Risograph. Spend your boldness in a single place; discipline everywhere else makes that one move land. (Coco Chanel's rule: before you leave, take one accessory off.)

Structure is information

Only use a device — numbers, dividers, labels — when it encodes something true.

Big 01 / 02 / 03 markers are fine if the content is genuinely a sequence. On a page that isn't, they're decoration pretending to be meaning. The Botanical page uses a two-column "sheet + catalog card" grid because the herbarium metaphor is real — the layout means something.

Motion that serves, and respects prefers-reduced-motion

One orchestrated moment beats scattered effects — and always give people the off switch.

The Midnight stars twinkle and the moon drifts, but both stop for anyone who's set "reduce motion" in their OS. Ambient atmosphere, not a fireworks show; excess animation is one of the strongest tells that a design was generated on autopilot.

The CSS-specificity trap

Two rules can quietly cancel each other; the more specific one wins.

A class like .section p can silently override .cta p and eat your spacing. It's the single most common way an AI-built page "looks off" for no obvious reason. Keep selectors flat, and run /code-review — it's tuned to catch exactly this.

The quality floor (never announced, always there)

Responsive to mobile, visible keyboard focus, readable contrast, honest alt text.

Distinctive and accessible aren't a trade — they're both just "done well." Every gallery page reflows to a phone, keeps focus visible, and respects reduced motion. Build to that floor without making a show of it.

The screenshot-critique loop

Build, look at a real screenshot, remove one thing, build again.

Claude designs far better when it can see its output. Use /layout or claude-in-chrome to shoot the page, read it back, and fix what the eye catches that the code review won't. This loop is where good becomes memorable.

Know them to avoid them

The three looks AI defaults to

The frontend-design skill calls these out by name. Each is legitimate for some brief — but they show up regardless of subject, which is the tell. If the brief leaves an axis free, don't spend that freedom on one of these.

Cream & terracotta serif

Near-#F4F1EA background, a high-contrast serif display, a single terracotta accent. Elegant — and everywhere.

Near-black + acid accent

Almost-black canvas with one bright acid-green or vermilion pop. Reads "modern" by reflex.

Broadsheet

Hairline rules, zero border-radius, dense newspaper columns. Fine for an editorial brief; a default for everything else.

Do it yourself

A five-step recipe you can run today

  1. Name the subject. One concrete thing, its audience, and the page's single job. Distinctive choices come from the subject's own world.
  2. Invoke the method. Ask Claude for the page and let frontend-design lead — or type /frontend-design. Have it draft a token system first: palette, type, layout, and one signature element.
  3. Critique the plan before building. "Which parts of this are the generic default for any page like this?" Revise those, then build from the revised plan.
  4. Look at it. Run /layout yourdomain.com, read the screenshots, fix what the eye catches. Run /code-review for the specificity traps.
  5. Systematize (optional). If it's becoming a reusable kit, push it to claude.ai/design with /design-login then /design-sync, so every future page designs against it.

Want the exact prompts? The Act page has copy-ready ones for adding a direction, running a critique, and shipping a skin. To see how the network measures up, open the design scorecard.