In the fast-evolving landscape of AI workflows and developer tooling, the intersection of output formats and user engagement is crucial. Recently, Thariq Shihipar, a prominent engineer and internal tools lead mikegingerich.com known for shipping developer workflows at scale, shared insightful thoughts around what he calls the “unreasonable effectiveness of HTML.” His commentary, especially framed through an X thread on Using Claude Code HTML for AI outputs, peels back assumptions about token costs and sheds light on how HTML artifacts, combined with smart templating and JSON-driven data swaps, can transform productivity and user engagement.
What are HTML artifacts in AI workflows?
To understand why Thariq advocates for HTML, we first have to place “HTML artifacts” in context. Within AI tools and prompt-driven pipelines, an artifact is the output format produced by the model. Historically, this has been plain text or sometimes Markdown. While Markdown—known for its simplicity and readability—has become a common choice for AI outputs, it carries several limitations. HTML artifacts push beyond Markdown by enabling richer, more interactive, and structurally clear outputs.
HTML artifacts are rendered snippets that allow:
- Semantic segmentation of content (headings, paragraphs, lists) Styling for scannability and accessibility Embedding interactive elements or pre-formatted code blocks Reusable templates powered by JSON or data injection
In Thariq’s words, these artifacts are not just a nicer way to display content—they’re a fundamentally richer output format that can dramatically improve engagement and workflow integration.
Engagement and scannability: The real drivers of better product work
One of the core themes in Thariq’s thread is that engagement and scannability of AI-generated content are key to unlocking true productivity benefits.
When developers or product teams consume AI outputs, they don’t want to wade through verbose or poorly structured text. They want to scan quickly and extract actionable insights. HTML’s native support for semantic tags such as
, , , and enables this effortless scanning. Markdown, while decent for plain text formatting, limits rich usability: Tables in Markdown are often clunky and inconsistent in rendering. Nested elements or interactive components are impossible or require hacks. Styling control is minimal or delegated to client applications. Thariq notes that when users can visually parse the output quickly, it drives product clarity and faster iteration cycles. This improved engagement ultimately means fewer follow-up clarifications, better team alignment, and a sharper feedback loop. Why token cost is a minor concern—latency is the real friction
Another misconception Thariq addresses is the obsession with token cost in AI output generation. This is especially relevant as some conversations about prompt design center on minimizing tokens to reduce expense.
According to Thariq’s experience, token cost, while nontrivial, is dwarfed by the impact of latency, meaning the time users wait to see the output. He argues:
“It’s not about squeezing token counts as the primary optimization — it’s about reducing latency and increasing the perceived speed and responsiveness of the AI workflow.”
HTML artifacts, although sometimes larger in token count compared to raw text or Markdown, do not meaningfully increase computational latency if generated and served efficiently.
Instead, designing smart HTML templates that can be easily swapped via JSON payloads helps keep APIs fast, reusable, and avoids repeated token-costly formatting patterns.
Reusable artifact templates with JSON data swaps
Thariq highlights a powerful technique tying all this together: template-structured HTML artifacts powered by JSON data injection.
This approach consists of:
Defining a generic HTML template with placeholders for content like titles, code snippets, tables, or bullet lists. Feeding the AI a structured JSON payload holding the specific data—like error messages, code diffs, or product specs—that gets swapped into the template. The AI’s job is less about formatting and more about computing the correct content for each slot.Benefits of this design pattern include:
- Consistent output formatting across multiple AI tasks Less output parsing complexity on the frontend or consuming tools Smoother versioning and maintenance of UI artifacts Enhanced clarity by separating data from presentation
The reusable templates concept is perfectly aligned with modern engineering best practices like separation of concerns and DRY (Don’t Repeat Yourself)—except applied to AI output.
How to apply Thariq Shihipar’s insights today
If you’re building or iterating on AI-powered developer tooling or product systems, integrating Thariq’s insights can elevate your workflows. Here are pragmatic steps:
- Move from Markdown to HTML artifacts: Encourage your AI pipeline to output well-structured HTML snippets instead of raw Markdown or plaintext. Design templates with semantic HTML: Build a library of reusable HTML templates featuring headings, lists, tables, and code blocks for common AI output types. Use JSON data injection: Separate your AI-generated content from presentation by producing JSON data that fills templates via simple swaps. Focus on latency reduction over token optimization: Prioritize architectures and prompt strategies that yield faster responses rather than solely minimizing token counts. Measure engagement and scannability: Collect feedback from your teams about how easily they can consume and act on AI outputs, then iterate template designs accordingly.
Conclusion
Thariq Shihipar’s reflections on the “unreasonable effectiveness of HTML” shed valuable light on how output format choices can make or break AI workflow productivity. Leveraging HTML artifacts enriched with semantic structure and driven by reusable templates plus JSON data provides a powerful pathway toward more engaging, scannable, and efficient AI-powered tools.
By shifting the conversation from token cost fixation toward latency optimization and better artifact design, teams can unlock deeper benefits from AI assistance. In short, embracing HTML as the lingua franca of AI outputs will help teams move faster, reduce friction, and write better products.
Cited inspiration from:
- Thariq Shihipar — expert in developer workflows and internal tools leadership Using Claude Code HTML — an example of AI integration with code snippets in HTML Relevant discussion threads on X (formerly Twitter) about AI output formatting