Getting Started with Kami: Stop Letting AI Documents Look Like Default Word Exports

The most underestimated problem in AI document generation is not content. It is delivery.

Modern models can draft research reports, resumes, proposals, presentations, and structured summaries from messy source material. Yet once the output lands on a page, it often has the same unmistakable default-AI look: white background, cold gray text, loose paragraphs, casually bolded headings, and SaaS-dashboard shadows scattered across cards. The content may be acceptable, but the layout has already drained away half of the reader's interest.

Kami is aimed at that gap. It is not simply another online document editor, and it is not just a bundle of pretty templates for an AI assistant. It turns the question of what a good document should look like into a set of constraints that an agent can execute reliably.

That distinction matters. Once AI-generated documents enter real delivery workflows, the problem is no longer only whether the model can write. The question is whether every generated deliverable can maintain a consistent aesthetic, density, reading rhythm, and print-ready structure. Content production has improved quickly; the quality of the page should not remain stuck at the level of a casual Word export.

The name Kami, from the Japanese word for paper, is direct. It points to the final layer that carries the work: the page.

The project by tw93 defines Kami as a document design system for AI. In its current form, it is an open-source Skill and template system that can be used by agents such as Claude Code, Codex, OpenCode, and Pi when they can read Skills. It can also be uploaded to Claude Desktop as kami.zip.

Its core idea is not, "Here are eight templates; pick one and edit it." A more accurate description is that Kami decomposes document delivery into three layers:

This makes Kami feel more like an operating system for document delivery than a folder of attractive screenshots. When an agent receives a task, it should not merely pour content into HTML. It should choose a route based on the document's purpose, language, audience, and output format.

If you ask for a research report, the agent should take the Long Doc path. If you ask for a company one-pager, it should use the One-Pager route. If you ask for a presentation, the default path is WeasyPrint HTML to PDF slides, with editable PPTX reserved for cases where that format is explicitly required.

That may sound like a small decision tree, but it is valuable agent engineering. It turns free-form generation into procedural delivery.

Kami's visual style is restrained: a warm paper background, ink-blue accents, serif type, compact line height, and very light shadows. It avoids the purple gradients, glassmorphism, and oversized cards common in many AI-generated pages.

The most interesting part of the system is not that it looks good. It is that it includes a clear negative space: things the agent should not do.

The page background is fixed at #f5f4ed rather than pure white. Pure white can feel harsh on screen and cheap in print. The warm paper tone is easier for long reading and gives the page a more editorial character.

The accent color is limited to a single ink blue, #1B365D, and it is used sparingly. Many generated documents collapse because the model adds too many colors: red, green, violet, and random gradients all competing for attention. Kami narrows the color space so emphasis appears only where it is needed.

The typography is equally opinionated. Chinese uses Canger Jinkai 02, English uses Charter, and Japanese uses YuMincho, with each language kept as close as possible to a single serif family. Hierarchy comes from size, spacing, and left-side rules rather than constant bolding. The result feels more like a publication than an admin panel.

There is also a small but telling engineering rule: tag backgrounds must use solid hex colors rather than rgba(). This is not design fussiness. In WeasyPrint PDF rendering, transparency can create a double-rectangle artifact where padding and glyph areas overlap. The rule shows that Kami is not only staging a pretty browser preview; it is accounting for the failure modes of PDF export.

Individually, these rules are simple. Together, they create the value. Agents do not fail because they have no design imagination. They fail because their design choices drift from one output to the next. Kami deliberately reduces the degrees of freedom so the document remains stable.

Document design systems are usually written for humans: brand guidelines, component libraries, design files, and examples. Kami is different because the reader is often an agent. It has to tell the agent how to decide, not just what to admire.

That means the instructions need to be operational. Which template is selected for which task? How should the agent handle long tables? When should a page break be forced? What should happen if the content overflows? Which export command is expected? What visual defects should be checked before delivery?

A human designer may absorb those rules through taste and experience. An agent needs them as explicit procedure. Kami's value is that it makes design intent executable.

Kami's default delivery path favors HTML rendered to PDF through WeasyPrint. That choice is practical. HTML and CSS give agents a familiar structure to generate and modify, while PDF remains the format many clients, teams, and readers expect for final delivery.

It also forces useful discipline. A document that only looks good in a browser preview is not finished. It must survive pagination, print dimensions, font fallback, table wrapping, image scaling, and page overflow. Kami's rules around spacing, tags, shadows, and fixed color values all point toward that export reality.

Editable PPTX still has a place, especially when a team needs to continue revising slides manually. But making PPTX the default can push the agent into a more fragile production path. For many static deliverables, HTML to PDF is easier to audit and reproduce.

Kami is most valuable when the document is part of a repeatable workflow rather than a one-off experiment. Research summaries, investment memos, product briefs, resumes, change logs, company profiles, and portfolio pages all benefit from consistent formatting. The more often the document type recurs, the more useful the constraints become.

It is also useful when the source material is messy. An agent can first organize the content, then apply a document route that gives the final output a stable editorial surface. That combination is the real promise: not just generated text, but a deliverable someone can read, print, forward, or archive.

Kami will not turn every agent into an art director. It will not solve weak source material, vague audience definition, or missing brand strategy. What it can do is prevent the most common failure: documents that contain reasonable content but look unfinished.

For AI documents to be taken seriously, the page has to carry the same care as the prose. Kami's approach is compelling because it treats visual quality as an executable system, not an afterthought. The agent still writes, organizes, and exports, but it does so inside a disciplined paper-like environment.

That is a better default than another blank white document with bold headings and a few generic cards. If AI is going to produce final deliverables, it needs a design system built for agents, not just a prompt that says "make it beautiful."