Skip to content

stack-diagram Turns a Sentence Into an Architecture Diagram

A new agent skill draws your stack with vendor-supplied logos and hands back a single shareable HTML file.

Draftship Editorial3 min read

An agent skill called stack-diagram takes one sentence describing your stack and returns an architecture diagram drawn with official logos. Its repository, caru-ini/stack-diagram, had 2,489 stars and 133 forks as of 2026-11-14, 13 days after it was created on 2026-11-01, according to the project's README on GitHub.

What the source confirms

The project is written in TypeScript and released under the MIT licence. You add it to an agent with npx skills add caru-ini/stack-diagram. The README lists Claude Code, Codex, Cursor and other agents that support skills as compatible. In the Claude apps, the alternative is to upload stack-diagram.zip from the repository's Releases page as a skill.

How the diagram gets made

The agent reads your description and fetches logos from svgl first, then devicon, then gilbarbara/logos, or from the brand's own repository. AWS, Google Cloud, Azure and Kubernetes icons come from each vendor's own icon package. The source says logos are never recoloured or redrawn. Lines are routed around text labels, and the agent renders the diagram and checks it before handing it over.

Every diagram is rendered in both light and dark themes. Tiles stay white so logos keep their own colours, while group labels switch to the brand's dark-theme version.

The default output is one HTML file with fonts and logos inlined, so it renders without a network connection. You can also get a PNG, a dependency-free React component or a Claude artifact. The PNGs and the React component are rendered from that same file.

The README's first example describes Next.js and a Hono API on Vercel, Postgres on Neon, Clerk for auth, Stripe for payments and GitHub Actions for deploys. Other examples cover a pnpm, Turborepo, Biome, Vitest and Lefthook monorepo, a docs chatbot that queries Supabase pgvector and answers with Claude, and layouts on AWS, Google Cloud and Azure using each vendor's icons.

After drawing, the agent offers once to post the diagram to a public gallery at stacks.caru.moe. It shows exactly what would become public (the light and dark PNGs, an English title, the tech stack and a display name) and posts only after you agree twice. The prompt itself is never sent. A GitHub login is optional and never displayed; it only lets you delete the post later. The gallery site's source is in the site/ folder.

Limits and open questions

The skill requires Python 3.9 or newer and network access for the logos; PNG export uses Playwright with Chromium. Logos are not bundled. They are downloaded at build time, and the README lists each source and licence, including devicon under MIT, gilbarbara/logos under CC0 and Lucide under ISC. Fonts ship under the SIL Open Font License 1.1. The README also notes that brand names and logos remain trademarks of their owners, and that Microsoft permits the Azure icons in architecture diagrams, training materials and documentation.

The source does not say how the agent behaves with services it cannot match to a logo, or how it resolves a request that is ambiguous about where a component runs. Beyond the agent checking its own render, it makes no accuracy claims. The star and fork figures are a single snapshot from one day.

What this suggests for builders

This likely matters most as a documentation chore killer. Keeping a current architecture picture in a README or wiki is the kind of task that slips; describing the system in a sentence you already know how to write is cheaper. It also suggests that agent skills, which are just packaged instructions plus assets, are becoming a distribution format in their own right. Whether the diagrams stay accurate as a project changes is up to whoever reruns the prompt.

Draftship, the AI app builder, works the other way round: you describe an app in plain language and it writes the code, builds it and hosts it, with a built-in database and user sign-in, and paid plans that add model choice, source code access and GitHub push. It does not draw diagrams. The shared habit is the starting point, since both begin with a sentence about what you want. If that sentence is doing the heavy lifting, writing effective prompts is worth a read.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.