Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Stop letting your agents write Markdown.
TL;DR
Theo says the real mistake is treating Markdown as the default output for agents — reacting to Thoric’s Claude Code post and Andrej Karpathy’s endorsement, he argues HTML is often better for specs, design explorations, PR explainers, and one-off tools because it’s easier to scan, compare, and interact with.
The strongest HTML use case is side-by-side thinking, not prettier documents — Theo likes prompts such as “generate six distinctly different onboarding approaches in a single HTML grid,” because one-pass comparison produces more variety than asking for one option at a time.
Interactive HTML artifacts can turn agent output into working interfaces — he highlights examples like adjustable design mockups, implementation plans with collapsible structure, and throwaway editors that export JSON or prompts back into Claude Code, calling this a powerful way to smooth messy data workflows.
He’s skeptical of some of the hype: HTML is noisier, more token-heavy, and often ugly — Theo pushes back on claims around image support, SVG quality, mobile responsiveness, and version control, and says novelty is part of why people read HTML outputs more than 100-line Markdown files today.
A big hidden lesson is that developers should write more disposable code — Theo says 70%+ of the code he “writes” with AI gets thrown away after one use, and that building one-off UIs to inspect CSVs, JSON blobs, or prompts can dramatically improve decisions and downstream software quality.
Karpathy’s bigger thesis resonated most: AI output will get more visual over time — Theo agrees with the idea that audio may be the preferred human input, but vision is the preferred output, with today’s jump from raw text to Markdown to HTML eventually leading toward richer simulations and interactive visual systems.
The Breakdown
Markdown Wars, Round Two
Theo opens by joking about the backlash to his earlier anti-Markdown video, then says he’s not anti-Markdown so much as anti-overusing it. What changed is that big names now agree: Thoric from the Claude Code team is posting about “the unreasonable effectiveness of HTML,” and Andrej Karpathy is publicly saying LLMs do well when you ask for HTML output.
A Sponsor Demo That Quietly Makes the Point
Before diving in, he shows CopilotKit as an example of why plain Markdown is a weak interface for agentic apps. Instead of spitting out a bad table, the app renders custom React components, interactive comparisons, and a second pass of model reasoning — basically a live argument for structured UI over text blobs.
Thor’s HTML Gallery: Useful, but Also a Little AI-Slop-y
Theo walks through Thoric’s article and its companion repo of 20 self-contained HTML files, immediately clocking that the gallery copy “feels agent-written.” He even runs text through Pangram Labs’ detector: the article reads as human-written, while the HTML examples come back as fully AI-generated — which becomes a funny recurring tension in the video.
Where HTML Actually Beats Markdown
The first convincing examples are comparative ones: three debounced-search implementations shown side by side, or four visual directions for an empty-state UI. Theo says this matches a trick he already uses — ask for four distinct options at once on different routes, because sequential prompts tend to converge on the same idea.
Theo’s Pushback: Images, Tokens, and the Novelty Effect
As Thor argues HTML has better information density, visual clarity, and shareability, Theo keeps checking the hype. Models still hallucinate image URLs and broken base64, many examples aren’t actually mobile-friendly, and once everyone starts spamming HTML links instead of giant Markdown files, the current “ooh, I’ll read this” novelty may wear off.
The Part He Really Buys: HTML as a Review and Verification Surface
Theo gets much more interested when the format becomes operational instead of decorative. He likes HTML PR explainers with better hierarchy, collapsible sections, and focus areas for reviewers, comparing the idea to Devin Review’s regrouped PR interface and saying alphabetical file-by-file review has always been dumb — now it’s just dumber at AI speed.
Reports, Research, and the Career-Hack Angle
One of his favorite uses is asking an agent to synthesize Slack, Jira, git history, or a codebase into a polished HTML weekly update or explainer. Theo practically lights up here: if you want to stand out at work, hand leadership a clean HTML summary of what shipped this week, because making people feel informed with one prompt is a very real career advantage.
Throwaway UIs, Anthropic Rant, and the Bigger Interface Shift
The final stretch gets more philosophical and more Theo. He argues developers need to get comfortable writing disposable code — custom one-off HTML editors, CSV visualizers, and prompt-export tools — then absolutely unloads on Anthropic’s artifact implementation as “terrible engineering” despite the model’s ability to generate useful interfaces. From there he lands on Karpathy’s broader point: AI output is moving from raw text to Markdown to HTML and eventually to richer visual systems, though Theo thinks there are important steps in between — and, naturally, jokes that “all roads lead to React.”
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
