@elliotchen100: Thariq from Anthropic’s viral HTML post hit 1.5M reads. On the surface, it’s about formatting aesthetics, but he’s actually outlining a brand-new workflow. Picking out the most technical points. First, HTML isn’t a document; it’s a throwaway editor. Take his example…

Thariq from Anthropic blew up his article on HTML yesterday, racking up 1.5 million reads. At first glance, it appears to be about formatting aesthetics, but he’s actually outlining a completely new workflow. Here are the most technically insightful points. First, HTML isn’t meant to be a document; it’s a throwaway editor. He gave a classic example: rearranging priorities across 30 Linear tickets. You ask Claude to generate an HTML page where each ticket becomes a draggable card sorted into four columns: Now, Next, Later, and Cut. At the bottom, you add a “Copy as Markdown” button to export the final order and paste it back into Claude Code. This pattern applies to any scenario where “describing things in text is difficult.” Adjusting animation easing curves, colors, cron jobs, or regex patterns—all are an order of magnitude more efficient visually than expressing them in plain text. Second, use an HTML network for specs instead of a single Markdown plan. His workflow goes like this: first, have Claude generate six proposals across different directions, laid out side-by-side in one HTML file for comparison. Pick one to dive deeper into for mockups and data flow diagrams. Finally, write the implementation plan. When starting a new session, feed the entire HTML network as context. The verification agent also reads through this HTML. This means the spec is no longer just a transitional artifact for humans; it serves as shared memory for multi-agent collaboration. Third, SVG is a severely underrated output format. Ask Claude to diagram the token bucket rate-limiting logic as an SVG flowchart, annotate key code snippets inline, and include a “Gotchas” section. One diagram explains a concept an order of magnitude more clearly than 200 lines of Markdown. Fourth, he honestly acknowledges the trade-offs. HTML generation takes two to four times longer than Markdown, and diffs are harder to review. However, given Opus 4.7’s 1M context window, those extra tokens become virtually negligible within the overall context. That’s the crux of his compromise. In a nutshell: Markdown operates on the implicit assumption that “a person will read it from start to finish.” HTML operates on the assumption that “people just want to scan the highlights and jump in to modify things.” The latter perfectly aligns with the reality of human-machine collaboration in the AI era.