HTML Is the New Markdown. Reviewing It Is the Missing Half.

By Marcin, founder of Markloop · Updated September 2026

In May 2026, Thariq Shihipar, an engineer on Anthropic's Claude Code team, published a post saying he had stopped writing Markdown files for almost everything and now has Claude Code write HTML instead. He shared 20 example HTML files with it: code reviews, architecture diagrams, research docs, presentations. Lenny's Newsletter and InfoQ both covered it. Chew Loong Nian then ran all 20 of Thariq's prompts through Claude Code in both formats and wrote up the results in Towards AI: HTML won 17. The three Markdown wins were outputs that stay inside the agent's loop and never get read by a person.

The argument

Shihipar's point is about the person reading the output. A long Markdown file gets skimmed. A page with layout, color, diagrams and collapsible sections gets read, and the reader stays connected to what the agent is doing. In his words, he uses HTML because "it helps me feel much more in the loop with Claude." His examples of good fits: specs, implementation plans, code review walkthroughs and data exploration.

Figure 1. The same plan as a Markdown file and as an HTML page.

How this looks day to day

The same month, Shihipar was a guest on Claire Vo's How I AI podcast (part of Lenny's Podcast Network), recorded at Anthropic's Code with Claude event. The episode covers how he works day to day. Three habits are easy to copy:

Weekly status updates as HTML. He sends his manager a weekly update as an HTML page instead of plain text, because it's "more likely to actually get read". If you send clients a weekly report, this is the easiest one to try.
A design system in one HTML file. He keeps his colors, typography, spacing and components in a single HTML file that he carries between projects. The agent reads it before writing anything, so new docs match his style.
Throwaway editing interfaces. When one section of a plan needs work, he asks Claude to build a small one-off UI just for editing that section, uses it, and deletes it.

The episode also has a chapter on adding comments and annotations to HTML plans. His approach: have Claude make the page commentable, with "a place to copy out my comments", then paste those back into Claude Code. That works when you are the only reviewer. When a manager or a client reads the page, their comments usually arrive in Slack or email, and someone retypes them into a prompt.

Where Markloop fits

I built Markloop for that second case. Markdown files live in the repo and get reviewed in pull requests. Google Docs have comments in the margin. An HTML page from an agent has neither. Paste it into Docs and the layout and diagrams break, and most of the people who need to read it will never open a pull request. So the feedback comes back as screenshots and email replies, and whoever owns the doc turns it into a prompt by hand.

Anthropic has started on this for Claude artifacts: people you share one with can now comment, with limits for reviewers outside your organization (details in who can comment on a Claude artifact).

What the reviewing half needs

Authoring got…	Reviewing still needs…
Agents that produce rich HTML in seconds	A place where anyone can open that HTML and comment on it, with nothing to install
Artifacts: live pages at a URL	Comments pinned to the sentence or element they are about, not "the third paragraph, I think"
Versions republished in place	Feedback tied to a version: what was addressed, what is still open
MCP: agents that can push work out	The reverse: comments going back into the agent with their context, not pasted by hand

The last row matters most to us. A comment like "§2 contradicts §3" helps the agent only if it arrives with the sentence it points to and the version it was made on. Figure 2 shows what that looks like.

One round, start to finish

Take the weekly status update. Your AI publishes it as an HTML page at one link. Your manager, Dana, opens the link and comments on one sentence. You ask your AI to pull the comments. Each one arrives as a short entry in comments.md: the target on the page, the quoted text, the text around it, the note and the version. The AI edits the page on your machine, publishes v2 at the same link and marks Dana's comment as addressed.

Figure 2. What your AI receives when it pulls a comment, and the next version at the same link. Field names match the comments.md file Markloop exports.

To try it with your own update or spec, ask your AI for the doc as a single HTML file, then say: "Publish this weekly update to Markloop and give me the share link." When Dana has replied, ask it to pull her comments and publish v2.

No setup needed to try it: download the HTML file your AI made and upload it to Markloop. Or connect Claude, ChatGPT, Cursor or Grok once (setup is a one-time step, shown in the app after signup) and ask your AI to publish. Watch the 2-minute walkthrough.

Sending it to a client? See how to write a client proposal with AI and get comments on one link. For who can see your link and where your data lives, see the FAQ.

Your agent already writes HTML.

Publish it at one link, let people comment on it, and have your AI pull the comments in for the next version.

Start your 14-day free trial Talk to us

14-day free trial · No credit card required · Reviewers always free