On May 8, 2026, Anthropic engineer Thariq Shihipar posted on X that HTML had become more useful than Markdown for agent output. His post arrived several months after Anthropic's own support team published the custom-visuals documentation stating that visuals shown to users in Claude and Cowork render as HTML, and the live-artifacts entry for Cowork, which describes the artifacts as interactive HTML pages users can edit, save, and rerun within a session.

The case for HTML output.

Shihipar's thread links to a companion gallery containing twenty HTML files generated by Claude Code. The examples include a pull-request review page with inline margin notes from the agent, a rate-limit diagram with sliders that recalculate the visualization when users adjust the inputs, a backlog view with a JSON export button, and a single-page weekly summary formatted for an executive recipient. None of those use cases is a one-off; an engineer on the Claude Code team posting twenty at once is documenting a pattern Anthropic's Console telemetry has presumably been showing for some time.

The closest historical analog comes from Apple's August 1987 release of HyperCard, an authoring environment Bill Atkinson designed and which Apple bundled free with every Macintosh sold for several years afterward. HyperCard let users build "stacks" for personal databases, technical reference works, and small applications using a scripting language called HyperTalk, without compiling anything. The stacks themselves carried persistent state, clickable buttons, and filtered views, which is the same intermediate slot HTML now occupies for users working with Claude Code.

This is the workflow argument behind Shihipar's preference. When an agent generates a plan as a Markdown file inside chat, the interface does not require the user to open the file before approving the agent's next action, and in practice the file is often skipped. When the same plan is generated as an HTML artifact, the file has to be loaded in a browser to be read at all, which inserts a review checkpoint that the chat loop would otherwise pass over.

Where Markdown still wins.

Knowledge work runs on a different clock. A note written in 2024 and reopened in 2026 has to survive a renamed folder, a broken outbound link, a forgotten tagging convention, and the editor's own change of mind about the original claim. Persistence under those conditions is what file format is for, and it is the question Shihipar's post does not address.

Inside an Obsidian vault, Markdown still wins on those terms. The Obsidian and Claude Code knowledge-base architecture I described earlier this year works because raw clippings stay read-only, the LLM-generated wiki pages keep visible source links, and every compiled claim can be traced back to a plain text file by filename and backlink. Andrej Karpathy's personal Markdown corpus, reported to run to roughly 100 articles and 400,000 words, became useful as model context for an operational reason: the files are greppable, diffable, and version-controlled in git, which lets the corpus be repaired one file at a time when a claim turns out to be wrong.

Vannevar Bush, writing in The Atlantic in July 1945, described his hypothetical Memex as a desk that stored "trails" between documents instead of a tool that simply displayed them. Obsidian's backlinks, source folders, and plain text files belong to that lineage. An HTML dashboard generated on demand by an agent does not, because the trail evaporates when the tab closes.

What Anthropic is already shipping.

Shihipar's post lands inside an existing pattern. Anthropic already ships Artifacts, the custom-visuals system documented in its support pages, live artifacts inside Cowork, downloadable .html exports, and the connector-aware pages I covered in last month's piece on Cowork. Cursor's version 3 release made the same bet on a similar timeline: the agent-orchestration UI shipped only after users had already been improvising ad-hoc dashboards inside chat for months.

The Anthropic product surface around Claude Code already contains most of the pieces. Anthropic has shipped Artifacts, the custom-visuals system documented in its support pages, and the live-artifacts feature inside Cowork. What Claude Code itself does not yet have is a first-class render pane or export command for the HTML files Shihipar's thread shows users already requesting through ordinary prompts. A future release adding those affordances would turn observed user behavior into supported behavior, and Anthropic's Console telemetry has presumably been showing that demand for some time.

The practical division for working users is narrower than the X debate suggests. HTML is the right output for files that will be opened, demonstrated to a colleague, scanned, and then discarded after a single review session, which describes most of the use cases in Shihipar's gallery. Markdown remains the right output in a different and smaller set of cases. A research note that will be reopened in 2027 after being filed in 2025, edited by hand without an authoring tool, and cited as the source of a later compiled claim has nothing to gain from being rendered in a browser at the moment of capture.

Whether Anthropic ships the export button in the next Claude Code release or the one after, the format division Shihipar's thread documented is already in operation, with users choosing HTML or Markdown on a prompt-by-prompt basis depending on what they intend to do with the file once the agent finishes.

Opinion

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]