HTML overtakes markdown as AI's preferred output format

May 18, 2026 · Lenny's Podcast
🎧 PodShort 359 min squeezed to 2 AI SprinklerAS AI / ML
Episode artwork
Tharik
Works on Claude Code at Anthropic
Lenny's Podcast
359 min squeezed to 2
Full episode from Lenny's Podcast
Quotable Moments

All of us are becoming these compute allocators now, right? And so you have to decide what is worthwhile spending the compute on.

TharikWorks on Claude Code at Anthropic
00:19

This is not even personal software. It's like micro-software on top of micro-software.

TharikWorks on Claude Code at Anthropic
00:50
Key Insights
  • Markdown became a popular way of interacting with AI agents, but the generated plans often became too long to read, leading to users (and even the speaker) stopping reading them, which is identified as a mistake.
  • HTML is emerging as the new markdown for AI agents, providing a richer communication medium that is both easier for humans to read and for agents to understand and produce high-quality outputs.
  • Plans, PRDs, and technical specifications still hold crucial importance, even with increasingly intelligent AI models, as they guide the user (now a 'compute allocator') on how to best spend the AI's processing power.
  • Users are evolving into 'compute allocators,' requiring them to make strategic decisions about how to utilize AI's compute resources, for example, balancing the cost of an 8-hour Claude run ($500) against the value of the output.
  • Utilizing HTML for AI interactions enables the creation of more interactive and visual communication, which significantly enhances user engagement and ultimately leads to higher quality outputs generated by the AI.
  • HTML allows users to generate custom, throwaway UI tools (dubbed 'micro-software') for specific, intricate tasks with AI models, facilitating more precise interaction and contributing to superior output quality.
  • Effective prompting with AI necessitates a delicate balance: providing sufficient context and requirements to achieve desired outcomes without excessively constraining the model, thereby allowing the AI optimal flexibility.
  • AI-generated HTML-based design systems and component libraries can function as 'living design systems,' promoting seamless collaboration across development, marketing, and design teams by ensuring consistency and reusability.
Metrics Mentioned
  • 8 hours (Duration Claude can run for to complete a task, influencing compute cost.)
  • 500 bucks (Estimated cost associated with Claude running for 8 hours.)
  • a thousand line markdown file (An example of a lengthy markdown file that became too long for users to read or manage effectively.)
  • 50 lines of code (Typical length of a markdown-based feature plan.)
  • 1 percent (The speaker's estimated amount of production code generated by AI, contrasted with a much larger volume of other AI-generated content like dashboards and interfaces.)
  • 25 components (The number of design components in the example app's design system.)

RevBots.ai View:

  • AI Sprinkler teams bolt on HTML outputs without integrating with core workflows.
  • ARM-stage orgs would treat HTML as dynamic artifacts in automated GTM systems.
  • Compute allocation decisions mirror SaaS Hoarder tool sprawl challenges.
🎧Full Episode:Lenny's Podcast →