You Need to Use HTML More, Again

In the agent era, the most undervalued output format is the oldest one. Markdown is a report you skim once. HTML is an interface you actually use. Here is the playbook — prompts, loops, templates, and a skill that makes every slide come out on-brand.

Nine words broke everyone's workflow

On May 8, 2026, an engineer on Anthropic's Claude Code team, Thariq Shihipar, posted nine words: "HTML is the new markdown." It did roughly 4.4 million views in sixteen hours. By the next morning, Simon Willison — who had spent three years typing some version of "just give me Markdown" into every prompt — had publicly reconsidered his default in a post he titled The Unreasonable Effectiveness of HTML.

That is not a fad. That is the industry quietly admitting something it forgot: the web platform was always the best canvas we had. We just stopped reaching for it because authoring HTML by hand was slow. Agents removed the slow part. What's left is the canvas — and it's enormous.

Markdown was the lingua franca of the chat era. HTML is the lingua franca of the agent era. The agent doesn't think in slides. It thinks in code. And code wants to be HTML.

This is the case for using HTML more — again. Not as nostalgia. As leverage.

Markdown is a report. HTML is an interface.

Here is Thariq's actual argument, and it is sharper than the meme. A long agent output written in Markdown is a report. You read it once. Past about line 15, you abandon it. The exact same content — same tokens, same information — rendered as a single self-contained HTML file becomes an interface: tabbed, filterable, searchable, navigable, sometimes interactive. And an interface gets used, not skimmed.

Same tokens. One you abandon at line 15. One you actually use.

Don't take the claim on faith. Flip the switch below — it renders one identical agent output two ways:

The data never changes. Only the container does. Markdown gives you a scroll. HTML gives you a tool. In the era where the marginal cost of generating either is the same prompt, choosing the scroll is leaving the entire web platform on the table:

The nuance everyone misses: Markdown for the loop. HTML for the deliverable. Use plaintext while you and the agent are thinking. Ship HTML the moment a human needs to read it.

HTML is how an agent shows its work — and asks what's next

This is the part that matters most for anyone running coding agents, and it's the most under-used move in the whole playbook.

When an agent finishes a chunk of work, the default is a wall of text: "I refactored the checkout flow, updated 7 files, and made some decisions." You scroll. You squint. You lose the thread.

Now imagine the agent ends every unit of work by emitting one self-contained HTML page instead: a review surface that shows the work and asks for the next step. What it changed (with collapsible diffs). The judgment calls it made. The open questions it needs your call on — as actual radio buttons. A proposed next-steps checklist. And a button that turns your selections into a clean instruction block you paste straight back to the agent.

That is the loop closing on itself. The agent's output isn't a status update — it's a control panel for the next turn. We built a working template you can steal:

• The Agent Work Review template →: — diffs, decisions, "I need your call" radios, and a one-click "compose guidance" that generates the paste-back instruction.

The best agent output isn't a summary of what happened. It's an interface for deciding what happens next.

Put the rule in your CLAUDE.md (or AGENTS.md) once and it compounds across every session:

## Output format
When you produce a report, audit, plan, dashboard, or work-review,
output a SINGLE self-contained HTML file (inline CSS + JS, no external
deps), not Markdown. End work-reviews with: what I did, decisions,
open questions as selectable options, and proposed next steps.

Keep that file short. Bloated instruction files get ignored — the rule gets lost in the noise. One crisp paragraph beats a page of nagging.

The prompts that produce great HTML

You don't need magic words. You need three habits, all visible in Anthropic's own frontend-aesthetics guidance: demand a single self-contained file, name the structure you want, and turn the ambition up.

1. Demand one self-contained file.

Output as a single self-contained HTML file with all CSS and JS inline
and no external dependencies. I want to open it directly in a browser.

2. Name the structure, not the pixels.

Build a dashboard for this CSV as one self-contained HTML file:
sticky sidebar nav, a summary stat row, collapsible sections,
client-side search and sort, print-friendly. No external dependencies.

3. Turn the ambition up — the highest-leverage modifier there is.

Create [X]. Include as many relevant features and interactions as
possible. Go beyond the basics to a fully-featured implementation.

That last line is the cheat code. The same model that hands you a flat page when you ask for "a dashboard" will hand you a genuinely impressive one when you tell it to go beyond the basics. We collected the whole set — prompts, loop recipes, and one-liners, each copy-to-clipboard — into a living playbook:

• Prompts & Loops Playbook →: — filter, search, copy. Bookmark it.

Save your brand as a skill — and every slide comes out aligned

Here is the move that turns "nice one-off" into "house style on tap."

In October 2025, Anthropic shipped Agent Skills, and made the format an open standard that December. A skill is just a folder — a SKILL.md plus any scripts and resources — that an agent discovers and loads only when it's relevant. The principles are composable, portable, efficient: build it once, and it works the same across the Claude apps, Claude Code, and the API.

The flagship example Anthropic published is a brand-guidelines skill. It carries the company's identity — colors Dark #141413, Orange #d97757, Blue #6a9bcc, Green #788c5d, fonts Poppins for headings and Lora for body — and applies them to artifacts automatically, right down to generating on-brand PowerPoint via python-pptx. (See the SKILL.md →.)

Sit with what that means. You stop re-describing your brand in every prompt. You save it once, as a skill — and from then on every deck, every report, every page your agent generates comes out aligned by default. The brand becomes ambient.

Describing your brand in every prompt is a tax. Saving it once as a skill is an asset. The asset compounds; the tax never stops.

This is the difference between an AI that can make a slide and an AI that makes your slide. Swap the tokens, and the entire deck re-skins in a blink — because the deck was never hard-coded, it was brand-driven the whole time. We built both halves so you can feel it:

• A brand-aligned HTML deck →: — arrow-key navigable, animated metrics, and a live brand-swap that re-skins all 8 slides instantly.
• Author a brand skill →: — edit colors and fonts, watch a branded artifact update live, and copy out the generated SKILL.md and brand.css.

A slide deck shouldn't be a file you reuse. It should be a function you call — with your brand as the argument.

The whole loop, end to end

Stitch the pieces together and you get a pipeline with no dead ends:

• Prompt: — describe the artifact; demand one self-contained HTML file.
• Build: — the agent ships HTML: layout, interactivity, and styling in one file.
• Review: — open it in the browser. It shows its work and asks for your call.
• Guide: — answer in the interface; paste the generated instruction back.
• Brand: — the saved skill makes it aligned without you asking.
• Publish: — HTML in, URL out. Now it's shareable.

That last step used to be the wall. Your agent generates a beautiful file and then… it lives on your laptop. You can't paste a file into Slack. You can't text it to a client. You can't open it on the projector without a song and dance.

That's the entire reason host-html exists: drop an HTML file (or have your agent POST it), get a live URL in seconds. It's the missing last mile — and it's exactly why this very article was also published as HTML, on the HTML host, as meta as it gets:

• Read this manifesto as a standalone HTML page →: — same words, rendered as an interface.

HOSTHTML_API_KEY="hh_YOUR_API_KEY"
curl -X POST https://suifgsvtcbrawzdhyuhk.supabase.co/functions/v1/publish \
  -H "Authorization: Bearer $HOSTHTML_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"html": "<h1>shipped by an agent</h1>"}'
# → { "url": "https://host-html.com/p/abc123", ... }

Steal the whole collection

Everything referenced above is sitting in one public collection, free to open, read the source, and remix. Six interactive, single-file HTML templates — the manifesto, the markdown-vs-HTML proof, the agent work-review, the branded deck, the prompts playbook, and the brand-skill authoring tool:

• Open the collection → /c/use-html-again

The bites

Steal these. They're built to be screenshotted:

HTML is the new markdown. — Thariq Shihipar, Anthropic Claude Code team, May 2026

The web platform never got worse. We just got a faster way to write it. The decade-long detour through rigid slide grids and disposable PDFs is ending, and the thing on the other side is the thing that was always there: a page. Living, responsive, interactive, shareable with a link.

Use HTML more. Again.

Drop your first HTML page → host-html