HTML overtakes Markdown as AI agents demand richer interfaces

May 18, 2026 · Lenny's Podcast
🎧 PodShort 36 min squeezed to 2 AI SprinklerAS AI / ML New
Episode artwork
Tharik
Works on Claude Code at Anthropic
Lenny's Podcast
36 min squeezed to 2
Full episode from Lenny's Podcast
Quotable Moments

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

Claire VoProduct Leader and AI Obsessive at How AI AI
02:108

I believe in you and trust you! Exclamation, exclamation, exclamation. I'm like, truly, I know you are capable and I believe in you to make these decisions.

Claire VoProduct Leader and AI Obsessive at How AI AI
03:18
Key Insights
  • While Markdown was popular for interacting with AI agents, increasing plan lengths made them unreadable, hindering oversight. However, plans, PRDs, and specs still matter for effective AI interaction.
  • AI models are essentially 'compute allocators,' meaning users must decide what's worthwhile to spend compute time on, as agents can incur significant costs for complex tasks.
  • HTML is emerging as the 'new Markdown' for AI agents, offering a richer, more visual communication medium compared to plain text, especially as agent interactions become more complex.
  • HTML allows agents to generate interactive, custom user interfaces and visual mockups (micro-software) within plans, enabling users to engage directly with and refine the agent's output for higher quality results.
  • When prompting AI, it's crucial to find a balance: provide enough information to get what you want, but avoid over-constraining the model, allowing it flexibility to provide broader context and solutions.
  • The ability to generate and consume content with AI at near-zero functional cost means rigid document structures (like PRDs and specs) are less necessary, allowing for more flexible and adaptable content creation.
  • HTML makes it possible to create 'living design systems' and 'living design repositories' directly within code, allowing for component visualization, variation testing, and real-time updates for developers, marketers, and designers.
  • The output of AI agents is increasingly focused on generating many 'tokens' for dashboards, custom interfaces, and interactive plans rather than just production code, allowing users to better define and refine their desired outcomes.
Metrics Mentioned
  • 8 hours (Duration Claude can run for a task, implying associated compute costs.)
  • 500 bucks (The cost Claude can spend on a task, highlighting that users are 'compute allocators'.)
  • 1000-line Markdown file (An example of an excessively long and unreadable plan generated by AI agents.)
  • 1% (The amount of tokens produced that go into production code, versus other tokens for dashboards, custom interfaces, etc.)

RevBots.ai View:

  • AI Sprinkler teams bolt on HTML generators without full workflow integration.
  • ARM-stage teams would treat AI as compute allocators with cost-aware protocols.
  • Visual planning artifacts could bridge gaps between RevOps and product teams.
🎧Full Episode:Lenny's Podcast →