As we move into an era where large language models act as our primary research assistants, the way we receive and process information has become a critical bottleneck. For years, the developer community has defaulted to Markdown as the standard output format. This convention was born out of necessity, constrained by the strict token limits of early high-performance models like GPT-4, where every character mattered. But as our context windows expand, capable of holding significantly more data than ever before, clinging to simple, flat-text Markdown might actually be limiting the practical utility of the answers we receive.

Simon Willison’s recent exploration of Claude Code highlights a compelling shift in how we should approach prompting: requesting HTML instead of Markdown. This transition is not just about bolding text or changing fonts; it is about unlocking the full capability of the web browser as an information interface. When you prompt an AI to output in HTML, you are asking for a structured, rich document that can natively house SVG diagrams, interactive widgets, and complex in-page navigation.

Consider the difference between reading a static PDF of a technical manual and navigating a well-designed, clickable knowledge base. Markdown acts as the static PDF, whereas HTML acts as the interactive application. By shifting the output format, you effectively transform the AI's response from a passive block of text into an active, functional tool. In his analysis, Willison points to practical examples, such as security exploit explanations, where instead of simply dumping code blocks, the AI can render interactive explanations, color-code specific findings by severity, and create visual margins that clarify complex logic in real-time.

For university students or researchers working across disciplines, this represents a massive upgrade in information synthesis. Instead of asking a model to simply "explain this code," you can refine your prompt to "explain this code using an HTML artifact with interactive diagrams." This simple change in the instruction layer allows the AI to organize information in a way that respects the inherent complexity of the data, rather than flattening everything into a rigid, text-heavy stream that is often difficult to parse.

While Markdown remains essential for documentation that needs to be portable and plain, HTML is clearly the superior choice for high-density, analytical tasks. It turns the AI from a mere text generator into a document engineer. As we continue to refine how we interact with these systems, the key skill won't just be what we ask, but how we direct the model to structure the results. By prioritizing HTML, we can transform every AI response into a bespoke, interactive experience tailored to the specific problem at hand, effectively lowering the cognitive load required to understand difficult concepts.

As we move into an era where large language models act as our primary research assistants, the way we receive and process information has become a critical bottleneck. For years, the developer community has defaulted to Markdown as the standard output format. This convention was born out of necessity, constrained by the strict token limits of early high-performance models like GPT-4, where every character mattered. But as our context windows expand, capable of holding significantly more data than ever before, clinging to simple, flat-text Markdown might actually be limiting the practical utility of the answers we receive.

Simon Willison’s recent exploration of Claude Code highlights a compelling shift in how we should approach prompting: requesting HTML instead of Markdown. This transition is not just about bolding text or changing fonts; it is about unlocking the full capability of the web browser as an information interface. When you prompt an AI to output in HTML, you are asking for a structured, rich document that can natively house SVG diagrams, interactive widgets, and complex in-page navigation.

Consider the difference between reading a static PDF of a technical manual and navigating a well-designed, clickable knowledge base. Markdown acts as the static PDF, whereas HTML acts as the interactive application. By shifting the output format, you effectively transform the AI's response from a passive block of text into an active, functional tool. In his analysis, Willison points to practical examples, such as security exploit explanations, where instead of simply dumping code blocks, the AI can render interactive explanations, color-code specific findings by severity, and create visual margins that clarify complex logic in real-time.

For university students or researchers working across disciplines, this represents a massive upgrade in information synthesis. Instead of asking a model to simply "explain this code," you can refine your prompt to "explain this code using an HTML artifact with interactive diagrams." This simple change in the instruction layer allows the AI to organize information in a way that respects the inherent complexity of the data, rather than flattening everything into a rigid, text-heavy stream that is often difficult to parse.

While Markdown remains essential for documentation that needs to be portable and plain, HTML is clearly the superior choice for high-density, analytical tasks. It turns the AI from a mere text generator into a document engineer. As we continue to refine how we interact with these systems, the key skill won't just be what we ask, but how we direct the model to structure the results. By prioritizing HTML, we can transform every AI response into a bespoke, interactive experience tailored to the specific problem at hand, effectively lowering the cognitive load required to understand difficult concepts.