AI 写出来的技术文档动辄几千字，全在终端里滚，没人愿意看。md2html 让 AI 自动把这些 Markdown 转成带侧边栏目录、图表、时间线、卡片和警告框的 HTML 页面，一个文件就能发给团队看。

https://t.co/c8nFrxzUNB https://t.co/gu12APZFdU

haidang1810/md2html

Source: https://github.com/haidang1810/md2html

md2html

Your AI writes docs. md2html turns them into pages people actually read. A portable skill for Claude Code, Codex, Antigravity — or any AI agent — that turns long-form Markdown (plans, specs, system designs, RFCs, runbooks, postmortems, brainstorms, notes) into a single self-contained HTML page with sidebar TOC, Mermaid diagrams, step timelines, callouts, comparison cards, and a light/dark Claude-orange theme.

The problem

Your AI just produced 2,000 words of system design. Beautiful logic, careful tradeoffs, three architecture options compared — and you can’t bring yourself to read past section 2 because it’s a monospace wall of text in a terminal scrollback. Same goes for the brainstorm doc you asked it for last week, the migration plan from this morning, the postmortem draft you’ve been ignoring. Nobody on the team will click your plan.md link either.

What this skill does

You type /md2html anydoc.md. Your agent reads the file, opens template.html + components.md from this repo, and writes one polished HTML file next to the source. No build step, no server, no Node, no Python.

Works on anything long-form your AI hands you:

• Plans — multi-week roadmaps, migration steps, rollout strategies
• Specs — feature specs, API specs, data-model designs
• System designs — architecture docs, RFC-style proposals
• Runbooks — operational procedures, incident response steps
• Postmortems — incident reviews with timeline + root cause + action items
• Brainstorms — option lists with rationale, idea dumps to organize
• Notes — meeting notes, research summaries, anything else markdown

It’s not a Markdown-to-HTML converter. It’s an analyzer. The agent decides which sections become Mermaid diagrams, which become numbered step cards, which become pros-cons tables, which become collapsible “deep dive” panels. The result feels designed, not converted.

Why it’s worth your ~/.claude/skills slot

• Reads like a product page, not a spec. Sidebar TOC with scroll-spy, anchor links on every heading, copy-to-clipboard on every code block, a scroll progress bar at the top. Things you’d build into a real docs site.
• Diagrams instead of paragraphs. Three-hop flows automatically become Mermaid flowchart blocks. Trade-off discussions become side-by-side comparison cards. The agent decides — you don’t write any of it.
• One file. Email it. Drop it in Slack. Open it on a plane. Self-contained HTML with embedded CSS and theme JS. The only network request is the Mermaid CDN, and you can inline that too if you care.
• Zero install for your users. No npm install. No Docker. The whole skill is three markdown/HTML files. Anyone with an AI agent and git clone can use it in under a minute.
• Portable across agents. Works the same way in Claude Code, Codex CLI, Antigravity, Cursor, Continue.dev — anywhere an agent can read a file and write a file. No agent-specific code.
• Source language → output language, automatically. Vietnamese spec? Vietnamese UI. Chinese plan? 目录 instead of Contents. Eight languages built into the label table; the agent fills in any other language by translation.
• Production-grade UX, not a weekend hack. WCAG AA contrast, ≥ 40 px touch targets, prefers-reduced-motion, focus-visible rings, mobile TOC drawer with backdrop and ESC-to-close, skip-to-content link, full print stylesheet. Reviewed by a UI/UX engineer who didn’t pull punches.

Install

Claude Code

git clone https://github.com/haidang1810/md2html ~/.claude/skills/md2html

Reload your session and run:

/md2html plan.md

→ outputs plan.html next to the source.

Codex CLI

git clone https://github.com/haidang1810/md2html ~/.config/md2html
mkdir -p ~/.codex/prompts
ln -s ~/.config/md2html/SKILL.md ~/.codex/prompts/md2html.md

(Codex’s prompts directory has shifted across versions — adjust the symlink target if needed.)

Antigravity

git clone https://github.com/haidang1810/md2html ~/projects/md2html

Open Antigravity → Settings → Custom Agents → New, paste the contents of SKILL.md, and give the agent file-read access to ~/projects/md2html/.

Any other AI agent

If the agent supports a system prompt or custom instructions AND can read/write local files, it can run this skill. Clone the repo to a stable path, paste SKILL.md into the agent’s custom instructions, done.

Usage

/md2html plan.md                    # writes plan.html
/md2html plan.md --out docs/x.html  # custom output path
/md2html                            # prompts for a file

The agent picks components based on what’s actually in the doc:

Multi-language

UI labels follow the source. The label table covers English, Vietnamese, Chinese (中文), Japanese (日本語), Korean (한국어), Spanish, French, German out of the box. Anything else, the agent translates equivalently and sets <html lang="..."> plus the --rec-label CSS variable. RTL languages are LTR-only for now.

Demo

See examples/example-plan.md (Vietnamese source) → examples/example-plan.html (rendered). The screenshots at the top of this README are from a longer test spec — a game economy design doc.

Customize the theme

All visual tokens are CSS variables at the top of template.html. Change accent color, surface, fonts, radii by editing the :root and [data-theme="dark"] blocks. Don’t rename classes — components.md references them.

:root {
  --accent: #D97757;    /* Claude orange — change me */
  --bg: #FAFAF7;
  /* … */
}

Update

cd ~/.claude/skills/md2html && git pull

Uninstall

rm -rf ~/.claude/skills/md2html

What’s in the box

md2html/
├── SKILL.md         # Instructions the agent loads on /md2html
├── template.html    # HTML skeleton: CSS + Mermaid + theme JS
├── components.md    # Component catalog the agent assembles from
├── examples/        # Reference .md → .html pair (gold standard)
├── docs/            # Screenshots for this README
├── LICENSE
└── README.md

Roadmap

• Inline Mermaid bundle (true offline mode)
• More themes (forest green, midnight blue, paper)
• Gantt / roadmap component
• Auto-detect images in source and inline them
• RTL language support (Arabic, Hebrew)

PRs welcome — especially translation tables for additional languages and theme variants.

License

MIT © 2026 haidang1810