Wink Pings

AI Keeps Drawing Diagrams Crooked, So This Open-Source Tool Takes the Pen Away from the Model

With a single /blueprint command, Claude gives you back an animated Victorian-style patent diagram, with parts popping out one by one by number. The visual style is just the shell; what the author really solves is another problem: how to keep diagrams generated by large models from coming out crooked.

I came across a Claude Code skill called Blueprint. Once installed, type /blueprint how does a cache work in the terminal, and it draws you an animated patent diagram: Victorian-era fine line art, numbered parts, and a plain-language narration line for each step. It's like someone tore the cache page out of Thing Explainer and turned it into a living manuscript.

Unlike tools that dump the entire architecture diagram in your face all at once, it draws step by step. In the cache diagram, step one draws the App and the database, noting that a query takes 120 ms; step two brings in the drawer by the door and explains why a Cache is needed; step three explains that the drawer is empty the first time, so you fetch once and keep a copy, and the next query returns in 2 ms. You understand what's missing in the previous step, and only then does the next part appear.

缓存图纸:App 每次都要走远路去数据库,于是门口放一个抽屉存最近的答案

Following the numbers does feel more intuitive than grinding through documentation. The site hosts 21 ready-made examples, covering the event loop, hash tables, rate-limiting buckets, CDNs, git merges, circuit breakers, and how a large model writes one token at a time. Each can be opened step by step, and you can copy the scene source code directly.

Up to this point, it's all product-level stuff. What's really worth discussing is behind the diagrams.

Many people have been burned by asking a large model to output SVG directly. Text models have basically no concept of two-dimensional space, so the diagrams they draw often have text over lines, boxes overlapping boxes, and arrows passing through parts—as if they have never seen what paper looks like. The model doesn't lack knowledge; it lacks the spatial sense to place that knowledge on a canvas.

The author of Blueprint didn't make the model force-write SVG. The model only outputs a tiny JS object: what parts there are, how they connect to each other, and how many steps they're divided into—that's it. The rest—layout, connections, animation, playback controls—is all handled by a zero-dependency runtime of about 65 KB. Parts are chosen from a kit of 28 pieces: drawers, cabinets, gears, gates, locks, clocks, queues, all ready-made.

There's also a self-check layer. Before handing over the diagram, the script lints the scene first, checking whether elements overlap, whether lines pass through text, and whether labels run off the page; then it launches Playwright to render key frames, lets the model take a look itself, and only sends it to you after confirming it isn't crooked.

There's a very honest line in the README: these validations matter much more than writing prompts. Let the model decide “what to draw,” let deterministic code manage “where to draw it,” and the hardest part of AI diagramming is basically cut open by this one stroke.

The author also clearly documents the visual sources of this system: the fine line art and numbered parts come from Victorian patent drawings, the plain-language labels learn from Randall Munroe's Thing Explainer, the approach of building a machine part by part comes from Bartosz Ciechanowski's interactive explanations, the idea of looking at concrete examples first and then abstract rules comes from Bret Victor's Ladder of Abstraction, and the skill packaging format references Kamran Ahmed's /stick-scene.

Install with one line: npx skills add intincrab/blueprint. Requires Node 18+, Playwright optional; installing it lets it render and check by itself, and also export GIF, MP4, and PNG, or package a diagram into a single HTML file to send to others. The project is under the MIT license, and the scenes you make belong to you.

Online examples: https://blueprint.intincrab.com

Open-source address: https://github.com/intincrab/blueprint

The visual style is the easiest part to copy. What's really worth borrowing is that division of responsibilities: semantics to the model, geometry to code, connected in the middle by a verifiable scene description. This structure isn't only useful for drawing diagrams; any time you have a model produce something with spatial constraints, it will probably come in handy.

发布时间: 2026-11-19 14:31