Can Claude make illustrations? Yes, in code

Yes, Claude can make illustrations, but it makes them in code. It writes SVG, HTML canvas or three.js that draws the picture in the browser; it does not generate raster images the way an image model does. That makes Claude good at clean, geometric, animated and interactive drawings, and weak at painterly or figurative art. Left to itself, each drawing comes out in a slightly different style. An illustration skill for Claude Code fixes that by giving it fixed rules to draw by.

What Claude is good at, and what it is not

Drawing in code plays to what a language model does well: structure, geometry and logic.

Good at: - Objects built from simple shapes: devices, tools, instruments, boxes, buildings in plan. - Diagrams, icons and spot illustrations with a clear idea behind them. - Motion and interaction. A lid that opens on click is a few lines of code, not a video. - Edits. "Make the line thinner", "use our blue", "open the drawer further" are changes to numbers, and nothing else in the drawing moves. - Fitting the page: a code drawing scales to any size, follows light and dark mode, and weighs little.

Not good at: - People, faces, hands, animals, cloth, anything soft or organic. - Texture, painterly light, photographic realism. - Judging its own work by eye. It writes the code blind unless it renders the result and looks at the picture. In Claude Code it can do that with a headless browser screenshot, but only if something tells it to.

Why the results are inconsistent without a system

Ask Claude for a drawing of a safe on Monday and a drawing of a key on Tuesday and you get two decent drawings that do not belong together. The camera angle differs. One has thick strokes and one has thin. One is flat colour, the other has a gradient. One has six parts, the other forty.

None of that is a model failure. A prompt like "draw a key" leaves dozens of decisions open (view, line weight, palette, level of detail, motion), and each run makes them afresh. For one hero image that is fine. For a product with empty states, a 404, onboarding steps and feature sections, it reads as clip art from five sources.

What an illustration skill for Claude Code does

A skill is a folder of instructions, a SKILL.md file plus any references and starter files, that Claude Code loads when it starts. When your request matches what the skill describes, Claude reads it and follows it. Skills install into ~/.claude/skills/ for every project, or into a project's own .claude/skills/ folder.

For illustration, a good skill closes the decisions a prompt leaves open:

  • The idea. How to choose an object that says the message without a caption.
  • The look. One camera, one line width, one palette, a limit on detail.
  • The motion. Which moves are allowed, and their timings.
  • The checks. Render the drawing, look at it, fix what is off.

Because the same rules apply every time, the tenth drawing matches the first.

A worked example: the Linejig skill

Linejig's free Line art skill draws interactive 3D line illustrations: one small object in line art in three.js, with one hero part that answers hover and click. It is tested in Claude Code and Codex CLI, and installs for Cursor, Gemini CLI, GitHub Copilot, Devin, OpenCode, Amp and goose too. Install it from any terminal (it needs Node 18 or later):

Code
npx linejig install

That copies it into the skills folder of each coding agent it finds, for Claude Code ~/.claude/skills/linejig; add --project to install it for one project only. Restart Claude Code, then run /skills and look for linejig in the list. There is a manual install without Node too.

Then ask in plain words and name the skill:

Code
make a figure for our 404 page with linejig

Here is what the skill tells Claude to do with that.

  1. Check the library first. If an existing figure already shows the thing literally, use it instead of drawing a new one.
  2. Think before drawing. Write three concepts, each as an object, its hero part, why it fits and its one move, then recommend one and wait for you to pick. A concept that needs a caption is dropped.
  3. Build on the shared engine. The figure file holds only the model, the order its parts assemble in, and what the hero does. Camera, line rendering, palette, intros, orbit, zoom and reduced motion all come from the kernel, so they cannot drift.
  4. Follow the drawing rules. A 2:1 axonometric camera, a handful of rounded solids, one line width of about 1 px at any size, five tones, one bright part. No text, labels or small detail inside the drawing.
  5. Render and look. Capture frames at the start, middle and end of the move, in light and dark, at desktop and phone width, and fix what is wrong before showing you.

The free skill draws in Line art. The Brand style, with the hero part in your brand colour, and the Dots style come in the Pro skill, Linejig Pro, which sits beside the free one.

Cartridge in 3D line art, with the cartridge picked out in orange
Cartridge Install it once.

What you get

A figure page that draws one object in three.js. It builds in when it appears (the parts stack into place, or the lines draw on), then waits. Visitors can drag to turn it, and it springs back to its composed view a couple of seconds after they let go. Hovering the hero part makes it respond the way the real part would, and a click gives it its full move. It follows light and dark mode, respects reduced motion by showing the final state, and can hand a click to your page as an action event, so a bell can open your help panel. The docs cover embedding and messages.

The figures in the library were drawn to the same rules, so a figure your agent draws sits next to them without looking out of place.

Alternatives, fairly

  • Image models such as Midjourney or Recraft are the right tool for raster art: painterly scenes, characters, textures, anything Claude cannot draw in code. The trade-offs are that the output is a picture rather than code, so editing a detail means regenerating, and keeping a set consistent takes style references or training.
  • Other illustration skills. illo makes editorial illustrations for articles, often with a recurring mascot. Colorfigures makes hand-inked, flat-coloured isometric figures of everyday machines as one HTML page you can turn and operate. Anatomy makes interactive isometric figures with many small precise parts in SVG and WebGL. Illos.dev sells animated illustrations with agent skills in several styles. Each has a different look; pick the one that matches your brand.
  • The page around the drawing. Anthropic's frontend-design skill works on type, colour and layout rather than illustration, and pairs with any of the above. There is more on that in why AI-built websites look the same.
  • An illustrator. For characters, a strong signature style, or a set that has to carry a whole brand, a person is still the best answer. A skill is good at extending a system; an illustrator is good at inventing one.

Next step

Install the free Line art skill from the skill page and ask for one drawing for a real screen in your product. Or start with the free figures, which embed with no account.

Published 21 November 2026.

Figures in this guide