AI Output Format Shake-Up: Markdown Falls Out of Favor as HTML Stages a Comeback

For the past two years, if you asked which format large language models preferred for delivering content, the answer was almost certAInly Markdown.

The reason was simple: Markdown is clean, free of extraneous formatting, and easy to copy into documents, knowledge bases, GitHub repositories, or even paste directly into a WeChat Official Account backend without causing major issues. In many ways, Markdown became widely recognized as the best Lightweight markup language of the AI era.

However, with the ARRival of the Agent era, Markdown's reputation has taken a sharp downturn. The old-fashioned HTML, long considered outdated, is now being enthusiastically recommended by developers. What triggered this shift? The answer lies in tools like anthropic's Claude Artifacts, OpenAI's Codex, and the recently popular Agent tools such as openclaw and Hemers — what they ultimately deliver is often not plain text. It could be an SVG animation, a working program, or even a complete video.

In such scenarios, Markdown, which typically renders only as plain text, simply falls short. Thariq Shihipar, a member of the Anthropic Claude Code team, noted in his own writing that claude should ouTPUt results in HTML rather than Markdown.

But is HTML truly superior to Markdown? Not necessarily.

The Old-Timer's Comeback Rests on Versatility

Markdown isn't particularly new. Conceived by John Gruber in 2004, it was positioned as a "text-to-HTML conversion tool" for web writers. In simple terms, it allowed writers to compose content in a format close to plain writing first, then convert it into HTML through a tool.

Markdown was never designed to replace HTML; its purpose was to lower the barrier to writing HTML. It served as an easier-to-use intermediary language, allowing authors to initially ignore the strict specifications required by HTML and focus on getting their content written smoothly, leaving the formatting to the conversion tool.

This is why Markdown gained popularity across blogs, forums, knowledge bases, developer documentation, and GitHub — it relied on an extremely simple writing syntax. The toolkit contained only headers, lists, blockquotes, bold text, and links — the most common formatting needs — sparing writers from having to worry about anything else.

HTML's history is much older. Tim Berners-Lee proposed the World Wide Web concept in 1989, and by the end of 1990, he had already DeFined several foundational concepts of the Web, including HTML, HTTP, and URL, while also writing the first browser, editor, and server. In effect, HTML was born alongside the internet itself.

Compared to Markdown, the cumbersome aspect of HTML is that writers must simultaneously focus on headings, paRAGraphs, links, images, tables, forms, navigation, buttons, and a host of other elements — potentially disrupting their train of thought. Yet HTML also renders far richer content. Its core purpose is to present a "mini web page," capable of displaying interACTive elements and animations that go far beyond Markdown's capabilities.

Using the Same Prompt, we asked an AI to create a personal resume with fictional details and output it in both Markdown and HTML. The Markdown version was predictably simple — the Information was complete, but if an HR professional received it, they might consider it unprofessional. The HTML version was far more refined; at the very least, it actually looked like a real resume.

We also compared a common product Comparison chart in both formats. The Markdown output was perfectly Standard — no unnecessary fluff, direct parameter information, all very clear. But in terms of detail richness, HTML again won decisively — consider elements like color choices or the most intuitive presentation of a recommendation score.

Does HTML win outright? I wouldn't be so sure. Using the same prompt, I asked the AI to create a startup founder's book recommendation list. With minimal intervention, I found the Markdown output clearer this time — book titles, summaries, and recommendation scores were all instantly legible. HTML delivered great visual effects, but it didn't allow me to quickly grasp which book I should start reading first.

Finally, let's examine the scenario of technical tutorials. Compared to previous cases, Markdown had an even clearer advantage here. Technical tutorials are inherently a Linear reading process — you browse from top to bottom, and Markdown's logic aligns perfectly with this. Code blocks are embedded directly within the steps, notes are properly indented, and everything looks clean. HTML, despite its high flexibility, had to revert to a simplified treatment in this context, and Markdown still APPeared more succinct.

From these real-world examples, it's clear that Markdown is not being completely crushed by HTML. Particularly in data collection and text organization tasks, Markdown delivers the clearest, simplest results. So why are experts recommending HTML? I believe it's because HTML can carry far more diverse content — in scenarios like resume building and product guides, the rich visual effects HTML brings are something Markdown simply cannot match.

Abandoning Markdown? It's Too Early to Say

Based on the experiences above, many might ask: if HTML looks this good, should everyone abandon Markdown and fully embrace HTML? Not really.

For most people, AI is a chat window — the output format doesn't matter much as long as you understand it. Markdown is perfectly adequate in this scenario and may even be more suitable precisely because of its simplicity. But when AI starts doing real work for you, Markdown might not shine as brightly.

The truth is, Markdown was designed from the start for "people who write." You write with it, submit it to a platform, and the platform translates those ## and ** into proper headings and bold text so users see the intended effect. More bluntly, Markdown inherently requires a "translator." HTML, on the other hand, opens directly as a finished product. The browser is its runtime environment, and every computer and phone in the world has one. HTML doesn't need anyone to translate it — it is the finished product itself.

ImAGIne asking an AI to make a PPT for you. If it outputs Markdown, what you get is not a PPT but a document describing what a PPT should look like. You still need to find a software that can read this "instruction manual" — like WPS or Microsoft Office — to see the actual slides. But if it outputs HTML, you open it and it's a complete, ready-to-use product.

This is precisely why nearly every expert recommends requesting HTML output from Agents — because that way, you receive an actual finished product, not a semi-finished one.

Furthermore, as mentioned, this is the Agent era. We must consider not only human-AI collaboration but also AI-to-AI collaboration. One Agent generates content, another receives and displays it, or pushes it directly to the user. In this pipeline, HTML can flow seamlessly — open and ready to use.

That said, Markdown's uses today are far more extensive than you might imagine. Consider the most classic example: during training, AI models ingest vast amounts of Markdown documents — project READMEs on GitHub, technical blogs, and various open-source documentation are mostly in Markdown format. As a result, AI is naturally familiar with Markdown's structure and comprehends it with the greatest ease. Many developers also prefer to share their techniques in Markdown format, making it convenient for other developers to feed directly into AI Models.

So, the fact that HTML is better suited to the Agent-era desire for a "perfect, finished product" does not mean Markdown should be phased out. Each has its appropriate scenarios and usage methods.

Markdown or HTML: That Is the Question

After all this discussion, as an ordinary user, how should you decide whether to request Markdown or HTML output from an Agent? It's actually quite simple. Just think about one thing first: is this result for your own eyes, or is it meant to be shared with others?

Take a travel itinerary as an example. You ask AI to organize a three-day Kyoto itinerary — your own plan, for your own use. Markdown is concise and direct, with all the information laid out clearly. In this case, if you insist on HTML output, you'll find the result barely differs from the Markdown version because HTML's design capabilities have no use here. The prompt becomes harder to write, and it's a completely unnecessary effort.

If you need to share it with others, especially in a relatively formal setting, HTML works much better. Consider a weekly work report: the Markdown version can only rely on EMOjis to differentiate project statuses, which might come across as too casual when sent to your boss. The HTML version uses green, yellow, and gray color indicators that are instantly readable, with progress bars clearly showing completion percentages — presentation handled impeccably.

Another rule of thumb: ask yourself whether you need the output to be beautifully formatted. If you don't care about layout at all and only want the raw information, Markdown is the right choice. But if you already have a vision in your mind — a specific layout, structure, color scheme, or animations — then HTML is the only way to go.

In summary, while HTML may be the output format heavily championed by experts for Agents, we still need to choose based on our specific use cases. Markdown and HTML each have their own strengths in their respective scenarios.