// the find
tw93/Kami
👩🚒 Good content deserves good paper.
Kami is an agent skill (for Claude Code, Codex, Cursor, etc.) that gives an LLM document templates and layout rules so it stops producing the same gray default-theme PDF every time. It covers eight document types plus landing pages and slides, rendered via WeasyPrint/python-pptx, and is aimed at anyone who regularly asks an agent to produce reports, resumes, or decks and wants them to look designed rather than default.
It actually has an eval suite — per-template graders for skill-fired, no-placeholders, content-fidelity, pdf-created — which is rare for agent-skill repos and tests the real failure mode (the model hallucinating placeholder content instead of using the brief). The Mermaid pipeline is a genuinely useful piece of engineering: it renders via beautiful-mermaid then re-themes and sanitizes the SVG to be WeasyPrint-safe, solving a real compatibility gap without pulling in a Node toolchain. Degradation is handled deliberately rather than ignored — Pygments is optional with a monochrome fallback, missing CJK fonts recover from a CDN, and the daily version check fails silently offline. The MCP server gives you a second, lighter entry point into the same rendering engine without loading the full skill prompt, so it isn't locked to one agent's plugin format.
The default Chinese font, TsangerJinKai02, is personal-use-only and requires a paid license for commercial use — bundled by default in an MIT-licensed tool, that's a licensing trap for anyone who ships client-facing output without reading the fine print. The landing-page template, the one output type meant to actually be deployed, ships example config only for Vercel (robots.txt, sitemap.xml, vercel.json) — the other seven templates are host-agnostic static files, but this one assumes a specific platform. The whole rendering pipeline leans on WeasyPrint and python-pptx, both of which carry real native dependency weight (Cairo/Pango for WeasyPrint), and there's no visible requirements.txt or dependency manifest in the tree for the dozen-plus Python scripts driving it. The README itself flags that the MCP server will render local file, HTTP, and HTTPS resources with the host process's own permissions — worth knowing before you wire it up to an agent that has broader network access than you intended.