Claude Code adds HTML artifacts and publish-to-link workflows in creator demos

• trq212's thread pushed a simple workflow into the open: ask Claude Code for HTML instead of Markdown, then use those pages for planning docs, code review, prototypes, and diagrams.
• According to LinusEkenstam's post, some creators had already been using HTML for context-building because it was easier for them to scan, while LinusEkenstam's reply said the format was also easier to articulate.
• a reposted publish flow showed the next step: generate an HTML artifact with Claude, tell it to publish, and get back a shareable URL.
• Official Anthropic surfaces now line up with that behavior: the Claude Code product page says the desktop app can preview servers and review visual diffs, while the Chrome beta docs describe browser testing and data extraction.

You can browse Thariq Shihipar's examples page, read Simon Willison's link post noting that Shihipar is on the Claude Code team, and inspect the Code with Claude event page, which framed the week around live demos and new capabilities. The weirdly practical bit is here.now, a hosting service built around the sentence “publish to www.here.now,” which turns an agent-made page into a public link.

Markdown suddenly looks underdressed for AI handoff docs. In trq212's main thread, the pitch was not better syntax, it was better artifacts: pages you can lay out, style, and skim like actual documents.

On the linked examples page, that breaks into five concrete categories:

• Exploration and planning: side-by-side approaches, visual directions, implementation plans.
• Code review and understanding: annotated diffs, module maps, narrative PR writeups.
• Design: design-token swatches, component review sheets, prompt-ready visual specs.
• Prototyping: animation sandboxes, clickable flows, interaction mocks.
• Illustrations and diagrams: inline SVG figures, flowcharts, annotated visuals.

That matches what LinusEkenstam described as a preexisting habit. He said he had been using HTML for context building since last year because it was easier for him to digest than .md, and hckmstrrahul's reply added that converting Markdown into HTML made previews better and made option-presentation easier.

The examples page leans hard on comparison views, which is probably why this idea spread so fast among design-adjacent builders. Instead of one long response, the page shows three code approaches next to each other, multiple design directions rendered live, and implementation plans with timelines, data-flow diagrams, and risk tables.

Simon Willison's writeup pulled out the same pattern in a PR-review prompt: ask Claude to render the actual diff, put annotations in the margins, and color-code findings by severity. That is a much more legible review surface than a wall of bullets.

A small supporting clue came from a repost about Claude Code preview, which pointed to desktop preview features that let users attach DOM elements or use a pencil tool for context. The official Claude Code page now says the desktop app can manage parallel tasks, review visual diffs, and preview servers from one place.

The missing piece was always distribution. the reposted demo reduced it to one prompt pattern: create HTML artifacts with Claude, tell Claude to publish, then get a URL back.

That makes these HTML pages behave less like local notes and more like disposable microsites. One person can generate a one-pager, prototype, gallery, or review doc, then hand over a live link instead of a file.

The service behind that flow, here.now, is blunt about the target user. Its homepage says, “Just tell your agent to publish to www.here.now,” then returns a URL at a here.now subdomain, no account required. The FAQ says it works with any AI agent that can make HTTP requests, including Claude Code, Cursor, Codex, and Amp.

This is still a community-led practice, but it is landing on top of official product surfaces that make HTML more useful than plain text. The Code with Claude conference page promised live demos of new capabilities, and ClaudeDevs' agenda post specifically listed sessions on “What’s new in Claude Code” and managed agents.

Anthropic's Chrome beta docs add another piece. They say Claude Code can connect to Chrome to test web apps, debug with console logs, automate form filling, and extract data from pages, with actions running in a visible browser window. For people already asking Claude to generate HTML, that closes a neat loop: make the page, inspect the page, revise the page.

The publish link itself has its own product logic, and that is where the workflow gets more interesting. According to the here.now FAQ, anonymous publishes expire after 24 hours, while free accounts make sites permanent.

The same FAQ says creators can publish under a custom subdomain or their own domain, password-protect pages, or require payment via stablecoins on the Tempo network. That turns a throwaway HTML artifact into something closer to a shareable client mock, paid report, or semi-private review room, without building a hosting stack first.