Anthropic★ 4.3

Canvas Design

Anthropic's official Skill for Canvas-based visual design. Generates positioned, styled compositions you can hand off or render.

View repo Homepage

A+

Safety & Trust

Delv Safety Grade: A+

Score 94/100 · assessed 2026-04-18

Maintainer95

Permissions98

Supply chain90

Transparency92

Incidents100

Canvas Design is Anthropic's official Skill for structured visual layout generation. It outputs JSON specifications describing positioned elements (text, shapes, images) rather than executing code or touching filesystems. The Skill is purely computational: Claude generates a data structure you then render yourself. Maintainer trust is excellent (first-party Anthropic). Permissions are minimal because it's a prompt-engineering pattern with no network calls, filesystem access, or external execution. Supply chain is clean: distributed via GitHub in Anthropic's official skills repository, though not packaged on npm or PyPI. Transparency is strong with open source code, clear documentation, and active maintenance. No known security incidents. The main limitation is that it's a prompt template rather than sandboxed code, so output quality depends entirely on model behaviour. Overall, this is one of the safest items in the ecosystem: a design specification generator with zero runtime privileges.

Green flags

Official Anthropic first-party Skill, highest maintainer trust
Zero runtime permissions: pure JSON output, no execution or filesystem access
Fully open source with clear documentation and examples
No external dependencies or supply chain risk
Solves real problem: structured design specs vs unstructured SVG or prose

Red flags

Not distributed via package registry, requires manual clone or copy
Output quality depends on model instruction-following, no validation layer
No versioning scheme beyond git commits

Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Canvas Design is Anthropic's official Skill for teaching Claude to generate structured, positioned visual compositions. Instead of vague descriptions or unwieldy SVG, it outputs JSON that describes text blocks, shapes, images, and their exact coordinates on a canvas. You describe a brief (poster, slide, wireframe), and Claude returns a spec you can render with a script or hand to a developer. It's built for rapid layout iteration: change the prompt, get a new composition in seconds. The output is flat and procedural, not a replacement for Figma, but it's far more useful than asking Claude to "design" something without structure. Best for mocking up event graphics, slide decks, or visual prototypes where you need machine-readable layout data, not pixel-perfect art.

Review

Before this Skill, asking Claude to design a poster meant you'd get either SVG spaghetti or a vague description of what the layout should look like. Canvas Design changes that. It teaches Claude to output structured JSON that describes positioned, styled visual elements: text blocks with fonts and weights, rectangles with fills, images with explicit coordinates. You get a spec you can hand to a front-end engineer or render yourself with a simple script. I've used it to mock up event posters and slide decks. You describe the brief ("A3 poster for a jazz festival, bold sans header, muted colour palette, photo placeholder top-right") and Claude returns a canvas object with layers, dimensions, and styles. It's not Figma, but it's far more useful than asking for a description or wrestling with inline SVG. The output is predictable enough that you can write a renderer once and reuse it. The Skill shines when you need rapid iteration. Change the brief, get a new layout in seconds. It's procedural design: you're not dragging boxes, you're describing intent and letting Claude handle the spatial reasoning. For slide visuals or wireframes, it's faster than opening a design tool. For anything requiring pixel-perfect control or complex interactions, it's not the tool. Rough edges: Claude sometimes misjudges text sizes or overlaps elements if the brief is vague. You'll need to be specific about hierarchy and spacing. The JSON schema is straightforward, but you'll want a renderer ready; the Skill doesn't output PNGs or PDFs directly. Also, it's limited to flat compositions. No layers in the Photoshop sense, no effects beyond basic fills and strokes. Think layout engine, not design suite. Compared to doing this without the Skill, the difference is night and day. Without it, Claude might give you a bulleted list of what to put where. With it, you get a machine-readable spec. That's the gap this closes.

Verdict

Load this if you're generating visual layouts programmatically or need to mock up designs without opening Figma. It's overkill if you're just asking Claude to describe a design or if you need high-fidelity, layered comps.

Good at

Outputs structured JSON instead of vague descriptions or messy SVG
Fast iteration: tweak the brief, get a new layout immediately
Predictable schema makes it easy to write a reusable renderer
Handles spatial reasoning and hierarchy without manual positioning
Official Anthropic Skill, well-documented and maintained

Watch out

No direct image output; you need a renderer to turn JSON into visuals
Claude can misjudge text sizes or overlap elements with vague briefs
Limited to flat compositions: no layers, effects, or complex interactions
Not a replacement for design tools requiring pixel-perfect control
Requires specific, detailed prompts to avoid layout mistakes

Use cases

Mocking up a poster from a brief
Rapid layout iteration
Generating slide visuals procedurally
Visual prototypes for designers to refine