HTML is the new Markdown: How Anthropic engineers are building with Claude Code | Thariq Shihipar

What It Covers

Thariq Shihipar, an Anthropic engineer on the Claude Code team, demonstrates how replacing markdown with HTML files transforms the planning and specification process — producing richer, more readable artifacts that keep builders engaged with agent outputs and ultimately improve the quality of what gets built.

Key Questions Answered

• •HTML over Markdown for specs: When Claude generates plans as HTML files instead of markdown, builders actually read them. Thousand-line markdown documents go unread and unedited, while HTML renders scrollable, visual, diagram-rich plans that pull readers in and prompt genuine engagement — directly raising the quality of what the agent builds next.
• •Compute allocator as the new PM role: Running Claude for eight hours costs roughly $500 in compute. The core skill of product management has shifted to deciding what is worth spending that compute on. Spec and planning phases are where those allocation decisions get made, making upfront clarity about goals more valuable than ever.
• •Micro-apps for editing specific plan sections: When one section of an HTML plan needs revision, ask Claude to build a throwaway custom UI specifically for editing that module — interactive fields, toggles, copy-out buttons — then paste the refined output back into the main plan. This targets precision edits without re-litigating the entire document in terminal.
• •Living design system as design.html: Store the design system as an HTML file rather than a markdown document. The file encodes colors, typography, spacing, and core components in a visually rendered, compressed format that Claude can reference directly in any new project folder, replacing design.md and eliminating ambiguity about component styling.
• •Prompt structure for HTML plans: Keep prompts short but include two specific anchors — the content type you need (code excerpts, mock-ups, logic diagrams) and an open-ended trust signal ("whatever is needed for maximum context"). Over-constraining with role prompts like "you are an expert planner" limits output quality more than under-specifying does.