A New Paradigm: HTML Replacing Markdown as the Preferred Choice for AI Collaboration
Recently, a discussion about AI output formats has been unfolding in the tech community. The core idea stems from an article by Thariq Shihipar, an engineer at Anthropic, titled “The Unreasonable Effectiveness of HTML.” He argues that in the age of AI Agents, having Large Language Models (LLMs) output HTML instead of Markdown is a more effective way for humans and AI to communicate.
This viewpoint was quickly shared and endorsed by Andrej Karpathy, former Director of AI at Tesla and a founding member of OpenAI. Karpathy added from a cognitive science perspective that about a third of the human brain’s processing power is dedicated to visual information, calling vision the “ten-lane highway” for information entering the brain. Therefore, AI output should not be limited to plain text but should evolve into visual content that includes colors, layouts, and interactivity. This discussion signals a potential shift in the human-computer interaction paradigm.
The Limitations of Markdown: As AI Outputs Grow in Complexity
Markdown, with its simplicity, lightweight nature, and ease of reading and writing, has long been the standard for developer documentation, notes, and simple content publishing. However, as AI capabilities advance rapidly, its limitations have become increasingly apparent.
Thariq Shihipar points out that when specs, plans, or reports generated by AI Agents grow dramatically in length and complexity, Markdown files exceeding a hundred lines become difficult for humans to read and digest. In AI-driven workflows, the creation and modification of these long documents are mostly handled by AI, diminishing Markdown’s original core advantage of being “easy to manually edit.” In terms of information density and readability, Markdown has become a constraint rather than a benefit.
The Power of HTML: From Text to Interactive Visual Interfaces
Compared to Markdown, HTML, combined with CSS and JavaScript, offers unparalleled expressive power. In his open-source project html-effectiveness, Thariq Shihipar demonstrates multiple examples proving that HTML can express almost any type of information. Its core advantages are reflected in the following areas:
- Exploring and Comparing Solutions: AI can generate multiple design proposals at once and present them in a grid layout on a single HTML page. This allows users to compare them side-by-side, intuitively weighing the pros and cons instead of switching back and forth between lengthy text descriptions.
- Code Review: AI can generate an HTML report containing highlighted code diffs, inline comments, flowcharts, and module dependency graphs. This visual interpretation of code is far clearer and more efficient than GitHub’s default diff view.
- UI Prototyping: HTML itself is an interface description language. AI can directly generate interactive web prototypes, even including controls like sliders and buttons, allowing users to adjust parameters in real-time and quickly validate design ideas.
- Generating Interactive Reports: AI can integrate information from multiple data sources like Slack, Git history, and databases to create HTML reports with charts, summaries, and interactive elements. The delivery and completion rates for these rich media reports are much higher than for plain text files.
- Creating Temporary Editors: For parameters that are difficult to describe and adjust with text, such as colors, animation curves, or regular expressions, you can instruct the AI to generate a disposable HTML editor on the fly. Users can make adjustments in the visual interface and then copy the result with a single click.
Furthermore, HTML files have a natural advantage in sharing. They can be uploaded to any web server (like Amazon S3) and opened in any device’s browser via a link, greatly enhancing collaboration efficiency.
Practical Considerations: Challenges and Trade-offs of HTML
Despite its promising future, the practical application of HTML still faces some challenges. The first is ecosystem barriers. Markdown is deeply integrated into existing developer toolchains, such as various documentation systems and code repository README files, and the migration costs are significant.
Second are maintenance costs and toolchain compatibility. Manually fixing flaws in AI-generated HTML is much more difficult than with Markdown. Additionally, tracking changes to HTML files with version control tools like Git is very hard, as the output of git diff is almost unreadable, creating new obstacles for code review.
In terms of performance, generating HTML requires more processing time and more tokens than generating Markdown—about 2 to 4 times slower. However, as top-tier models with million-token context windows become more common, this cost is becoming acceptable. A pragmatic view is that Markdown can serve as an exchange format between AI Agents or for intermediate process outputs, while HTML is better suited as the final deliverable for humans.
A Practical Guide: How to Start an HTML-Driven AI Workflow
Adopting HTML as an output format doesn’t require complex technical setup. As suggested by Karpathy and Shihipar, users can simply append a command to the end of their prompt, such as structure your response as HTML or make me an HTML file.
The core shift is in mindset: instead of just asking the AI to “write” an answer, ask it to “build” a visual, interactive tool or product to present the answer. Thariq Shihipar’s example gallery, html-effectiveness, provides nearly 20 examples covering nine major categories like planning, design, reporting, and learning, offering a rich reference for users.
As Shihipar says, switching to HTML makes him feel more deeply involved in the collaboration with AI than ever before. This isn’t about pursuing flashy visuals, but about enabling humans to truly comprehend the complex outputs of AI, thereby maintaining control in the decision-making process.
