Deep dive · product tour · ~30 minInternal talk · July 2026
WIP — draft

Claude Design.

Describe it, see it, ship it — and how to make everything it makes look like your brand, not like AI.

describe refine ship
2
Agenda

Three parts, half an hour

PartTopicNote
01What it isthe map — interface, inputs, the brand feature
02How to work itthe craft · example briefs · live demo
03From canvas to shippedsharing, exports, Claude Code handoff

Claude Design is a beta — all product claims are date-stamped July 2026 and sourced on the closing slide.

01

What it is

The map before the craft: the canvas, what it makes, what you can feed it — and the brand feature that sets it apart.

3
What it is · in one slide

Describe it — see it

  • Describe what you want → Claude builds a working first version → refine by chatting, commenting, or editing directly.
  • From Anthropic Labs — launched April 17, 2026, 1M+ users in week one, still officially a beta.
  • Built for people without a design background — the model does the layout, typography and color work.
  • Lives at claude.ai/design (web & desktop app — no mobile yet).
It builds
real HTML

Under the canvas, a frontier Claude model (Opus 4.7 at launch) writes actual working code — which is why outputs can be interactive, and why the Claude Code handoff in part 03 works.

4
What it is · the interface

Chat left, canvas right

CHAT "Build a one-pager for the Q3 launch…" briefs · structural changes "try a different approach" build refine CANVAS — the working design click an element → inline comment drag, resize, edit text directly sliders & controls spacing · color · layout

Two panes, one rule: the conversation drives structure, the canvas takes the detail work.

5
What it is · what it makes

Decks to working prototypes

Decks & one-pagers

Pitch decks, presentations, briefs — the classic "I need slides by Thursday."

Pages & collateral

Landing pages, social assets, campaign material — marketing without the queue.

Wireframes & mockups

Product screens to react to — before engineering spends a sprint on them.

Interactive prototypes

Clickable flows without writing code — stakeholders drive it, not squint at it.

Design explorations

"Show me four directions" — compare side by side instead of marrying idea #1.

Beyond slideware

Because it writes real code: prototypes with voice, video, shaders & 3D.

One tool, one skill to learn — the same describe → refine loop drives every output type.

6
What it is · starting points

Four ways to start

Describe it a text brief in chat — the blank-page start Upload a file DOCX · PPTX · XLSX — ugly deck in, designed out Capture a website yours as brand baseline · a competitor's as reference Point at a codebase reads your components → production-shaped ONE CANVAS working first version same describe → refine loop

Codebase tip from early users: attach the relevant subdirectory, not the whole monorepo — very large repos cause lag.

7
What it is · the headline feature

It knows your brand

  • Every project automatically inherits your organisation's design system — colors, typography, components. Nothing to configure per project.
  • Feed it from a GitHub repo, design files, or a local codebase — the /design-sync command in Claude Code pushes it in.
  • One org can hold several design systems — product brand, marketing brand, sub-brands.
  • It checks its own work — builds with your real components, then reviews the output against the system and corrects deviations before you see them.
  • Answers the classic complaint: "AI output never looks like us."
GitHub repo design files local codebase /design-sync Org design system colors · typography · components every project inherits it nothing to configure self-check
8
What it is · plans & usage

Included — but it shares your budget

QuestionAnswer — July 2026
Which plans?Pro, Max, Team, Enterprise — included, from ~$20/mo. No separate license.
What does it burn?The same usage pool as chat & Claude Code — no separate design allowance; optional extra usage.
What burns most?Complex projects on large codebases — attach less, spend less.
Enterprise?Admin must enable it in org settings first.
Where does it run?Web & desktop app only — no mobile yet.

Budget mindset: a design session spends the same tokens your coding sessions do — plan heavy design days like heavy coding days.

02

How to work it

The craft: the loop, the brief, the three refinement channels — with example prompts to steal, and a live build.

9
Craft · the loop

Judge iteration 4, not iteration 1

  • The first generation is a conversation opener, not the deliverable — real value shows up in iterations 2–4.
  • Start simple, then layer: core layout & content first; interactions, edge cases and polish as separate rounds.
  • Claude responds well to incremental requests — ten small asks beat one ten-part brief.

If you remember one thing from this part: don't judge generation one.

describe review refine repeat ×2–4 1 2 3 4 opener judge here
10
Craft · the brief

Brief it like a creative director

✗ Weak

"Make me a landing page for our new API."

→ a generic template with placeholder everything.

✓ Strong

"Build a landing page for our new Payments API, aimed at backend developers. Hero with a one-line tagline and a curl snippet, three feature cards with icons, a pricing section with two tiers."

→ reviewable in one shot.

The formula: who it's for · which sections · what goes in each. You're not designing — you're briefing, like you would an agency.

11
Craft · refinement channels

Three channels — pick the right one

Chat — structural rearrange · add a section · new direction feels like: talking to the designer Inline comment — targeted click the element, say what's wrong feels like: markup on a proof Direct edit — by hand drag · resize · fix a typo · sliders feels like: a slide editor the whole design one element one detail CANVAS hero section 💬 "brand color here" card button

The classic mistake: structural work via twenty inline comments — or asking chat to nudge one button. Match the channel to the size of the change.

12
Craft · feedback that works

Say it so it lands

In chat — structural

"Rearrange the dashboard: metrics in the top row, the chart below."

"Turn this one-pager into a 6-slide deck — one section per slide."

In a comment — targeted

"Tighten the spacing between form fields to 8px."

"Use the primary brand color here."

"Change this to a dropdown instead of radio buttons."

Vague feedback is the #1 failure mode — "this doesn't look right" gives the model nothing. Name the component, give a number when you have one.

13
Craft · exploring safely

Explore without losing work

  • The branch phrase: "Save what we have and try a completely different approach." Explore freely — the version you liked survives.
  • Ask for variations up front: "give me four visual directions" → compare side by side, then commit to one.
  • Say early if it must work on mobile — responsiveness is a brief-time requirement, not a final-polish fix.
  • Use Claude as an accessibility reviewer: "check contrast, font sizes and hierarchy" — it's genuinely good at this.

Treat the canvas like a cheap sketchbook, not a precious file — branching costs one sentence.

gen 1 gen 2 the one you liked ✓ saved — survives branch — one sentence "try a completely different approach" dir A dir B dir C dir D four directions — compare side by side, then commit
14
Craft · briefs to steal

Four briefs to steal

Deck

"A 8-slide pitch deck for [product] for non-technical execs: problem, solution, demo screenshot, market, pricing, roadmap, team, ask."

Dashboard

"An analytics dashboard: revenue, active users & churn as top-row metric cards, a monthly trend chart below, filterable by region and plan."

Onboarding flow

"A 4-screen mobile onboarding flow: welcome, permissions, profile setup, first-run tips — clickable, with a progress indicator."

Internal tool

"A request-intake form for the legal team: requester info, request type dropdown, conditional fields per type, urgency, file upload."

Same formula every time: audience · sections · content per section · one constraint. Copy, adapt, iterate.

15
Craft · live demo
Live demo

One-pager in five minutes

  • Brief from the previous slide, pasted verbatim → first generation.
  • One structural change in chat · one inline comment · one direct edit · one slider.
  • Watch the design system do its work — everything lands on-brand without asking.

Live The point is the method, not a finished artifact — five minutes, hard stop.

16
Craft · fallback

Fallback: recorded run

Placeholder — screenshots or recording of the five-minute build: the brief, generation one, and one example of each refinement channel (chat · comment · direct edit · slider).

Use only if the live build stalls. Narrate the captured before/after per channel, then move on.

17
Craft · pitfalls

Six pitfalls early users hit

Vague feedback"Doesn't look right" goes nowhere — name the component, give a number.
Judging gen 1The first version is an opener — the value is in iterations 2–4.
Pixel-perfectingDon't chase the last 10% here — export to Canva for final polish.
Whole monoreposLarge repos lag — attach the relevant subdirectory instead.
Vanishing commentsBeta bug: inline comments occasionally disappear — paste the text into chat.
Co-editingSimultaneous editing is basic — treat it as one driver + commenters, not Figma.
03

From canvas to shipped

Share it, export it, or hand it to Claude Code — pick by where the design needs to end up.

18
Ship · share & collaborate

Share for feedback

  • Org-scoped sharing: private by default → view-only, comment, or edit access via link.
  • The workflow that works: share with comment access — stakeholders leave inline notes exactly where they mean them.
  • Resolve comments in one pass — each note is a targeted refinement Claude can act on directly.
  • Co-editing caveat (beta): one driver, many commenters — simultaneous editing isn't Figma yet.

Stakeholder comments become refinement input — the review round and the iteration round are the same round.

Your design — private by default share a link view-only comment edit 💬 stakeholders leave inline notes exactly where they mean them each note = a refinement Claude acts on beta: one driver, many commenters — not Figma yet
19
Ship · exports

Export by destination

Your canvas Export ↗ — button upper right PPTX · PDF a meeting — your deck culture Standalone HTML clickable prototype, one file Org URL feedback — view · comment · edit Canva the last 10%, in marketing's tool .zip raw HTML & assets, anywhere Claude Code handoff a repo — next slide

Export button, upper right. Rule of thumb: URL for feedback, file for meetings, handoff for engineers.

20
Ship · Claude Code handoff

The Claude Code round trip

Your repo components · design files Claude Design the canvas /design-sync IN — your design system handoff bundle → Claude Code OUT — implementation against the real repo
  • In: /design-sync (in Claude Code) pushes your design system from a GitHub repo, design files, or a local codebase.
  • Out: the handoff bundle carries the finished design to Claude Code for implementation against the real repo.
  • Because Design built real HTML with your components all along, handoff is a translation, not a rewrite.

Design explores, Code ships — and the shared design system keeps both honest.

21
Close · recap

Describe → refine → ship

01 Describe

Brief like a creative director: audience · sections · content · one constraint. Start from text, a file, a site, or a codebase.

02 Refine

Chat for structure, comments for tweaks, hands for the rest. Judge iteration 4, not 1 — and branch with one sentence.

03 Ship

URL for feedback, file for meetings, handoff for engineers — the design system rides along everywhere.

Homework: steal one brief from slide 14 and run it on a real task this week — on your plan, it's already included.

22
Close · sources

Sources — all claims dated July 2026

Claude Design is a beta — features, limits and plan details move. Re-check the Help Center before quoting externally. The 1M-users-in-week-one figure follows the series fact pass for deck 01 (July 2026).