The markdown-viewer/skills repository has fundamentally restructured how AI coding agents handle diagram generation in Markdown. Instead of scattering engine-specific skills across the ecosystem, the project now ships a single, self-contained package called documd-visuals. This consolidation follows the Agent Skills format standard and works across Claude Code, Codex, Cursor, and GitHub Copilot. The package contains 242 verified examples spanning 183 scenarios and 31 goal domains, all gated by automated rendering checks.
One Package Replaces Fourteen Fragmented Skills
The repository previously shipped 14 engine- and domain-named skills, which have now been rewritten and merged into documd-visuals. The old uml, cloud, network, security, vega, graphviz, and architecture skills no longer exist as standalone packages. Their content migrated into structured directories within the single skill: engines/, examples/, goals/, and styles/. There are no compatibility stubs left behind. The CLI discovers skills by locating a SKILL.md file, and a stub directory without one would be silently ignored rather than installed. Users must reinstall to pick up the new package.
Five Engines, Thirty-One Goal Domains, Nine Themes
The skill package recommends exactly five rendering engines: draw-uml 1.5.2 for UML and architecture diagrams, vega 6.4.0 and vega-lite 6.4.3 for data-transform charts, echarts 6.1.0 for dashboard-grade visualizations, @antv/infographic 0.2.20 for template-driven infographics, and bare HTML for system architecture layouts. Anything outside these five is explicitly not recommended. The SKILL.md router matches user intent to one of 31 goal domains grouped into four meta-clusters: data and metrics, process and systems, infrastructure and governance, and knowledge and expression. Every figure is colored from one of nine themes defined in styles/, and theme drift is caught by scripts/check-all.mjs before the build passes.
Installation Is a One-Liner, But the CLI Is Separate
Installing the skill is as simple as running npx skills add markdown-viewer/skills, which works with multiple AI coding agents simultaneously. However, the documd CLI for exporting rendered documents to DOCX, PDF, EPUB, or image formats is a completely separate tool. The Agent Skills format has no install-hook field, so the skill installer cannot trigger npm install for the CLI. The correct package name is @markdown-viewer/documd β the bare name documd on npm is an unrelated package called 'markdown object notation.' The npm release carries the same version as the extension, so an npx run uses the build documented in the skill.
Key Takeaways
- The documd-visuals package is self-contained: it never cites research/ or scripts/, and the build gate fails if it does. Every file inside the package is reachable from SKILL.md, enforced by validate-skills.mjs.
- The skill package is licensed CC-BY-4.0, covering examples and redistribution. The documd CLI and rendering engines (@markdown-viewer/draw-uml, @markdown-viewer/drawio2svg) are separate works under GPL-3.0-only.
- Automated gates include theme drift checks, contrast validation, block render verification, and a full render pass via scripts/verify-examples.mjs. All must exit 0 for a change to land.
- File budgets are strict: SKILL.md is capped at 300 lines and 12 KB, goal docs at 200 lines, and example files at a 120-line target with a 200-line hard limit for verbose Vega specs.
The Bottom Line
This is what mature agent tooling looks like. Instead of throwing 14 half-baked skills at LLMs and hoping the model picks the right one, markdown-viewer built a single router with 242 verified examples and CI gates that catch theme drift and render failures before they ever reach an agent's context window. The Agent Skills ecosystem needed this kind of rigor. If you are building AI coding agents and your diagramming stack is a pile of incompatible skill directories, delete them all and run npx skills add markdown-viewer/skills. One package, five engines, thirty-one domains, zero ambiguity.