Why Markdown is dead for AI coding agents (and what I built instead)

If you've been using AI coding agents like Claude Code, Cursor, or Antigravity recently, you already know they are a massive productivity multiplier. But as I started asking my agents to handle more complex refactors and architecture planning, I hit a massive friction point.

The Wall of Text.

When you ask an agent to plan out a complex feature, it typically spits out a 500-line Markdown response right into your terminal or a narrow IDE chat window.

Let's be honest: when an agent emits a thousand-line markdown plan, nobody reads it carefully. You skim it, assume the AI knows what it's doing, and hit "proceed." And that is exactly when careful review matters most.

Markdown is great for simple Q&A, but it's terrible for architectural review.

So, I decided to fix it.

Meet html-planner ◐

I built a lightweight tool called html-planner. It intercepts the markdown plans generated by your AI agents and serves them locally as a live, interactive web app.

Instead of reading a wall of text in your terminal, the agent opens a browser window where you can interact with the plan directly.

Here is what you can do with it:

1. Visually review the architecture
   If your agent writes a mermaid block in its plan, html-planner renders it as a live SVG diagram. You get actual architecture you can read, not ASCII art.

2. Approve or reject individual steps
   Every step in the agent's plan renders as a card with an approve and reject button. The agent cannot start writing code until you've signed off on the steps.

3. Answer questions before code is written
   If the agent needs clarification, it doesn't spam your CLI. It generates interactive question cards (radio buttons, text inputs) on the page. You answer them, and the agent reads your decisions back before proceeding.

4. Inline PR-style comments
   You can highlight any sentence in the agent's proposed plan, click Comment, and leave a note. The agent reads your edits and adjusts its approach.

How to use it
It integrates seamlessly out-of-the-box with tools like Claude Code, Cursor, and Antigravity by dropping a skill file into their respective hidden directories (e.g., .claude/skills/).

You can install it in your current project without cloning any repos or setting up databases. Just run:

bash

npx html-planner@latest init .
From then on, just tell your agent to "use html-planner" when you ask it to plan a complex task.

Try it out
I built this primarily to scratch my own itch, but I figured other heavy AI-agent users might find it useful too.

You can play around with the live interactive playground here: html-planner.com

Let me know what you think in the comments, or if you have any feature requests for your specific agent workflow!