Kami gives AI-generated documents a real design system

Most AI-generated documents are still trapped in the same aesthetic failure: they are technically correct, but visually forgettable. The content may be useful, yet the result still feels like a default export from a tool that does not really care about reading experience. That is why Kami stood out to me. It is not trying to generate random polished-looking pages. It is trying to give AI-generated documents a coherent design system.

Kami is an open-source project from tw93 built around a simple idea: good content deserves good paper. In practice, that means giving agents one constraint language for six document formats, including one-pagers, long documents, letters, portfolios, resumes, and slides. Instead of hoping the model improvises its way into decent typography and spacing, Kami encodes opinions about layout, hierarchy, colors, tone, and page rhythm so the output feels composed rather than assembled.

That framing matters a lot. A big problem with AI-generated documents is not that models cannot produce text. They obviously can. The real problem is that generation and presentation are often treated as separate concerns. You get decent words first, then some vague hope that another layer will make them look professional. Kami takes the opposite view. Presentation is part of the product from the beginning, not a cosmetic pass added at the end.

What I like most is how intentionally narrow the design language is. Kami is not a kitchen-sink template pack. It uses a warm parchment canvas, a single ink-blue accent, serif typography, and strict rules around shadows, spacing, and visual hierarchy. That might sound restrictive, but that is exactly why it feels useful. Good systems reduce decision noise. When you are asking an agent to turn research into a deck or a brief into a one-pager, fewer aesthetic degrees of freedom often produce a more reliable result.

The repo also feels product-minded in a way many design-adjacent AI projects do not. It is built around real outputs people actually need to ship:

• startup one-pagers
• investor or research-style long docs
• formal letters
• project portfolios
• resumes
• presentation slides

That sounds obvious, but it is an important distinction. A lot of open-source AI design work still optimizes for demos. Kami feels closer to a practical publishing layer for knowledge work. It is less about showing that an agent can make something visually interesting once, and more about making useful document formats look consistently better across repeated runs.

I also think the multilingual angle is underrated. English and Chinese are treated as first-class paths, while Japanese is supported with a best-effort workflow and visual QA. That is a healthier way to design language support than pretending every language path is equally solved. It tells me the project cares about output quality enough to be explicit about where it is strong and where users should verify more carefully.

Another reason I find Kami interesting is that it exposes a deeper truth about AI product design: quality often comes from constraints, not from raw flexibility. Builders sometimes assume the ideal AI tool should be able to produce anything in any style. In reality, the tools people trust most usually have a strong point of view. Kami has one. It wants documents to feel editorial, calm, and readable. That opinionated boundary is what makes it more credible as a workflow component.

There is also a nice meta lesson here for anyone building AI-assisted products. If the output quality keeps drifting, you probably do not just need a better model. You may need a tighter system. Kami is essentially that argument turned into an open-source artifact. It says that better prompts are not enough; you need reusable rules that encode taste and structure in a way agents can follow reliably.

Of course, it is not trying to be universal. If you need loud visual branding, highly customized layouts, or a broad design marketplace with endless stylistic variation, Kami may feel too opinionated. But I think that tradeoff is a strength. Focus is what turns a neat repo into a useful tool.

My takeaway: Kami is one of the more thoughtful open-source projects I have seen lately because it treats AI-generated documents as a design systems problem, not just a content generation problem. For builders who care about the last mile between 'the model wrote something' and 'this looks ready to share,' that is a very worthwhile direction.

GitHub: https://github.com/tw93/Kami