Descriptions:
Thariq Shihipar, an engineer on Anthropic’s Claude Code team, sat down at the “Code with Claude” developer conference in San Francisco to explain a workflow shift gaining traction inside the Claude Code team: HTML files replacing Markdown as the preferred medium for AI agent specifications, planning documents, and PRDs.
The reasoning is practical. As Claude Code agents have grown to run for an hour or more — spanning models like Opus 4.5 and 4.7 — Markdown planning documents have ballooned to thousands of lines that nobody reads. HTML offers real visual structure: scrollable layouts, tabs, mockups, and formatted type interfaces that are genuinely easier to scan, edit, and reason about. Claude handles HTML generation natively without resorting to the awkward ASCII art approximations common in plain-text plans. Shihipar notes that the richer format also enables a single document to hold multiple logical sections — a PRD in one tab, a tech spec in another — something Markdown can only approximate with separate files.
Host Claraveo, a product leader and creator of the “How I AI” podcast, draws out additional insights on spec design: using type interfaces as the essential skeleton of any feature, writing explicit test and validation criteria as the second bookend, and why Anthropic’s newly announced Outcomes feature represents a meaningful shift toward goal-oriented agent tasking. The conversation is an unusually candid look at how even the team building Claude Code thinks about staying in the loop as agents grow more autonomous.
📺 Source: How I AI · Published May 18, 2026
🏷️ Format: Interview
