Forty-two diagram types replace a blank canvas with rules
Diagram Design gives an AI coding agent a visual system rather than another drawing surface. The README lists 42 types, covering familiar architecture, sequence, state, timeline, and data-model work alongside Sankey, Wardley, journey, deployment, and database-schema layouts. Each type comes with light, dark, and editorial variants. A request becomes one self-contained HTML file with inline SVG, which you can open locally or place in documentation.
The useful constraint is deletion. The skill asks the agent to choose a type, limit node density, reserve one accent for the important part, and load only the relevant reference. That should produce a more coherent result than prompting an agent to improvise SVG from scratch. It also means you are accepting this project's taste. Our 15.5 MB checkout contained 474 files and about 42,413 lines of source, much of it devoted to examples, references, import rules, and validators.
Imports preserve meaning while replacing the original layout
The import paths accept draw.io containers, Mermaid files or Markdown fences, and Excalidraw JSON scenes. Diagram Design extracts components, relationships, grouping, and direction, then redraws them in its own grid and palette. You choose a destination size, output format, audience, and one of three detail levels. The resulting fidelity ledger lists what the agent kept, collapsed, or dropped, which is the right place to catch a missing service or decision branch.
Import rewrites the source rather than round-tripping it. The README explicitly says source coordinates, fonts, colors, and renderer layout do not carry over. A carefully placed Excalidraw sketch will come back as the skill's composition, and a Mermaid diagram will lose its automatic geometry. That bargain makes sense when 42 shared layouts are the goal. In the 15.5 MB checkout we reviewed, the tool was a bad fit for preserving pixel positions or producing the same picture from a text-only diff without an agent.
Brand onboarding changes files and can stop an unattended job
Onboarding can fetch a public site, map its background, text, accent, and font choices into semantic roles, then show a proposed change before writing it. Saved profiles let an agency reuse that result across projects. The first-use guard is deliberate: if the style guide still has its default palette, the skill pauses and asks whether to onboard, enter tokens, or continue with the default.
That prompt creates a real automation limit. Open issue 262 describes an unattended multi-diagram run that reached the style gate with nobody available to answer. The proposed noninteractive path was still an issue when we checked. The 37 MB installed footprint does not make this a background-safe batch tool by itself. If diagrams run from cron or CI, settle the profile and first-run behavior before putting the command on a schedule.
What happened when we ran it
Our sandbox installed commit 57148ac in 20 seconds. It added 35 Python packages and occupied 37 MB on disk. The build completed successfully in another 11 seconds inside an unprivileged Debian container with 3 CPUs, 8 GB of RAM, Python 3.12, and no secrets. Pip-audit found 0 known vulnerabilities. Those figures cover repository setup and build work, not diagram quality or export speed.
The test stage was skipped because our harness found no tests script or target. That does not mean the repository lacks checks. Its README describes custom Python validators for skins, geometry, imports, motion, accessibility, documentation links, and rendered output, plus CI on Linux, Windows, and macOS. The narrower finding is that commit 57148ac did not expose the conventional test entry point our sandbox expected, so adopters should map the project's checker commands into their own release job.
Static files are simple; raster export adds a browser
Ordinary output is static HTML and SVG with no app server. Motion is optional, the complete static frame remains available, and reduced-motion behavior is specified. PNG export is heavier: the documented route installs Playwright and Chromium, waits for fonts, and captures the diagram. Generated pages request Google Fonts by default. The privacy document says removing the font link or opening the file offline uses system fallbacks instead.
That font choice affects both privacy and layout. A browser opening the normal file contacts Google, while offline rendering changes the typefaces and can change text fit. The repository's rendered-layout checker therefore has separate default and real-font modes. Our 11-second build did not exercise a production export workload, so teams shipping slide decks should test their own longest labels, chosen scale, and offline policy instead of treating a successful build as visual approval.
Current activity is strong, with two practical gaps still open
GitHub showed 43,031 stars, an October 1, 2026 push, and 27 open issues and pull requests combined. The API split that queue into 15 issues and 12 pull requests. There was no latest GitHub release and the tags endpoint returned no tags, so update tracking follows repository commits and marketplace behavior rather than a conventional release page. The active queue and recent push point to maintained work, while the missing release channel deserves an explicit pinning policy.
Two open defects can decide adoption. Issue 254 says Arabic and other right-to-left output lacks documented bidi, font, and mirrored-layout rules. Issue 161 reports 2.86:1 contrast for the default accent on the paper color, even though the templates include accessible names and the onboarding flow checks text contrast. Neither gap was exercised by our 3-CPU sandbox build. Diagram Design is a strong fit for agent-written English technical documentation, provided a human reviews the generated file. RTL publishing and hands-off batch generation should wait for their named gaps to close.