For the last year or so, many of us have built a simple habit around AI:
ask a question, wait for a long Markdown answer.
Headings, bullet points, tables, bold text, code blocks. It looks organized, and it is genuinely better than plain text. But lately I have been feeling that asking AI only for Markdown is a bit like buying a color TV and using it as a radio.
Markdown is great. It is light, clean, portable, and perfect for notes and documents.
The problem is that many tasks do not really end with “text.” What we actually want is something we can look at, share, present, inspect, or publish: a report, a dashboard, a one-page brief, a set of cards, a lightweight slide deck, or even a page with filtering and interaction.
That is why Andrej Karpathy’s recent suggestion caught my attention. He wrote that you can append something like structure your response as HTML to the end of your prompt, then open the generated file in a browser. He also mentioned having success asking LLMs to present output as slideshows.
It sounds like a small trick. I think the direction is much bigger.
AI output should not always stop at “an answer.” Increasingly, it can become temporary software, a temporary page, or a temporary interface.
Markdown is not weak. It is just often an intermediate artifact.
Markdown has obvious value. Developers like it. Writers like it. Knowledge bases like it.
But from an ordinary user’s point of view, Markdown often stops one step before the thing is actually useful.
Ask AI to prepare a market research summary, and it gives you a Markdown table. Then you still need to copy it into a document, adjust the layout, decide what should become cards, and clean up the hierarchy.
Ask AI for a competitor analysis, and it gives you headings and bullet points. Then you still need to decide what belongs on the first page, what should become a comparison matrix, and what should be hidden in the appendix.
Ask AI to summarize a meeting, and it gives you structured text. But what you may really want to send to your team is a clean action page with priorities, owners, deadlines, and risks.
That is the awkward part of Markdown. It is more structured than plain text, but it is not always close to the final deliverable.
HTML is different.
HTML is natively understood by the browser. It supports layout, color, components, links, images, tables, buttons, and interaction. You do not need to translate it into another tool before it becomes useful. It can be the deliverable.
That is what makes Karpathy’s tip interesting. It is not merely a prompting trick. It is a reminder that the browser is already the most widely available runtime for content.
Asking for HTML turns an answer into an interface
The same request feels very different depending on whether the output is Markdown or HTML.
Imagine asking AI to analyze a batch of user feedback.
The Markdown version might look like this:
- users complain most about slow loading
- the second major issue is failed payments
- the third is confusing onboarding
- performance should be fixed first
That is readable, but flat.
The HTML version can become something else:
- key metrics at the top
- issue categories as cards
- raw user quotes on the left
- fix priorities on the right
- next-week action items at the bottom
Add a little JavaScript, and it can filter by tag, switch views, or expand original quotes.
At that point, AI is no longer giving you only “an answer.” It is giving you a small tool.
This matters a lot for creators. Blog posts, newsletters, short videos, internal briefs, and social posts are not only text. They are information interfaces. Previously, we asked AI to write the draft and then did the formatting ourselves. A better workflow may be to ask AI to organize the information as a visual page first, then extract the article, images, or video script from that page.
In other words, do not use AI only as a writer. Let it act as a partial designer and frontend builder too.
This does not mean everyone needs to learn frontend development. The value is almost the opposite. Ordinary users do not need to write HTML by hand. They only need to describe what kind of page they want, and AI can produce a first version they can actually inspect.
Why this feels newly important now
If Karpathy were the only signal, this would still be a useful trick.
But the signal is wider than that.
On Product Hunt, display.dev describes itself as a way to publish agent-generated HTML behind company auth. That positioning quietly assumes something important: agents will generate more HTML pages, and companies will need safe, controlled ways to share them internally.
On GitHub, ppt-master is gaining attention for generating native editable PPTX files from documents. PPTX is not HTML, of course, but the underlying direction is similar: AI is no longer only producing text. It is getting closer to final-format deliverables.
Product Hunt also has Jotform Claude App, which lets users build, edit, and analyze forms directly inside Claude. A form is not an answer. It is an operational interface.
Put these signals together, and the direction feels clear to me.
The competition in AI output is moving from “who writes a longer answer” to “who gets closer to a usable result.”
That is just common sense. Users are not using AI because they want to read AI essays. They use it because they want to skip a step, deliver faster, and turn a fuzzy thought into something that can be shared, run, or collaborated on.
Markdown is an intermediate layer. HTML, PPTX, forms, dashboards, and workflows are closer to outcomes.
When you should ask AI for HTML
I would start with four kinds of tasks.
1. Reports and summaries
Weekly updates, industry briefs, competitor analysis, investment notes, project summaries.
When these are written only in Markdown, they often become long documents. As an HTML page, they can have a summary strip, key cards, data tables, risk flags, and action items. Readers do not have to read from top to bottom to understand what matters.
2. Presentations and explainers
Karpathy mentioned slideshows, and I think that is especially practical.
You can ask AI to generate an HTML slide deck where each screen contains one point, a simple layout, and speaker notes. It will not replace a polished presentation, but it is good enough for internal sharing, course drafts, and short video storyboards.
For early thinking, an HTML slideshow can be much faster than opening PowerPoint and dragging boxes around.
3. Comparisons and structured decisions
Tool selection, pricing comparisons, product evaluations, vendor reviews.
Markdown tables become ugly as soon as they get large, especially on mobile. HTML can support collapsible sections, filters, sticky headers, and color-coded priorities. Once information density increases, HTML has a real advantage.
4. Visual cards for personal knowledge bases
Not every note should remain a wall of text.
A book’s key ideas, a technology learning path, or a paper’s structure can often be turned into a card-based page. The page does not need to be public. Even as a local browser file, it can be much easier to scan and revisit.
HTML has real pitfalls too
I like this direction, but it is not magic.
The first pitfall is security. If AI-generated HTML includes external scripts, suspicious links, or complicated JavaScript, do not casually open or publish it. For internal use, I would usually ask for a single-file page with no external scripts or remote assets.
The second pitfall is taste. AI loves producing pages that look like generic SaaS landing pages: huge gradients, giant cards, heavy shadows, vaguely premium but very formulaic. My advice is to ask for structure first and tune the visual style later. Do not chase “stunning” too early.
The third pitfall is maintenance. Markdown is excellent for long-term storage because it is stable and readable. HTML is better for display and delivery, but it should not always be the only source of truth. For important work, keep a Markdown or plain-text source as well.
The fourth pitfall is fact-checking. A polished page makes people trust the content more easily. That can be dangerous. AI-generated HTML reports still need checked citations, links, and data. Never let layout do the job of evidence.
How I would use the trick today
If you want to try it, do not start with a huge task.
Try adding this to a normal request:
Please turn your answer into a single-file HTML page.
Requirements:
1. Do not use external JavaScript or CSS.
2. Put five key conclusions at the top.
3. Use cards for the detailed analysis.
4. Put an action checklist at the bottom.
5. Keep the visual style clean and readable in a browser.
For report-style tasks, add:
Please put important data into tables and use color to distinguish high, medium, and low priority.
For presentation tasks, use:
Please generate a single-file HTML slideshow. Each slide should contain only one main idea and support left/right arrow navigation.
These prompts are not fancy. But they often produce something much closer to a deliverable than “please answer in Markdown.”
Final thought
I do not think Markdown is going away. It is too useful, especially for writing, knowledge bases, and version control.
But I do think we should stop defaulting to one reflex: asking AI for a big block of Markdown every single time.
Sometimes the better request is simply: give me an HTML page.
The deeper shift is that AI is no longer only an answer machine inside a chat box. It can generate temporary interfaces, documents, pages, and small tools.
And the browser is the easiest runtime most people already have.
So I am not surprised that Karpathy’s suggestion resonated. It is not a gimmick. It exposes a layer many people were ignoring:
if AI can already organize information, why force it to deliver only a text draft?
References
- Andrej Karpathy on asking LLMs to structure responses as HTML: https://x.com/karpathy/status/2053872850101285137
- Product Hunt: display.dev, “Publish agent-generated HTML behind company auth”: https://www.producthunt.com/posts/display-dev
- GitHub: hugohe3/ppt-master, AI-generated native editable PPTX: https://github.com/hugohe3/ppt-master
- Product Hunt: Jotform Claude App, “Build, edit, and analyze forms directly in Claude”: https://www.producthunt.com/posts/jotform-claude-app