Using Claude Code: The Unreasonable Effectiveness of HTML

2026-05-08 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, short

Summary

Thariq Shihipar of Anthropic's Claude Code team advocates for using HTML over Markdown as an output format when prompting large language models like Claude. This approach, detailed with examples on a dedicated site, allows for richer explanations incorporating SVG diagrams, interactive widgets, and in-page navigation, enhancing information clarity and user experience. Historically, Markdown was preferred due to token efficiency with models like GPT-4's 8,192 token limit. However, the ability of HTML to create dynamic and visually engaging content, such as color-coded PR reviews or detailed exploit explanations, suggests a reevaluation of output format defaults. An example demonstrates GPT-5.5 generating an HTML explanation for a Linux security exploit from obfuscated Python code, showcasing the potential for interactive and well-styled technical documentation.

Key takeaway

For Machine Learning Engineers or NLP Engineers generating technical explanations or code reviews, reconsider defaulting to Markdown. Your outputs can be significantly enhanced by requesting HTML, enabling interactive elements, SVG diagrams, and advanced styling. This shift allows for more navigable and clearer documentation, especially for complex topics like security exploits or intricate code logic, ultimately improving comprehension and reducing review time.

Key insights

HTML offers richer, more interactive LLM outputs than Markdown, improving clarity and navigation for complex information.

Principles

• Prioritize output richness over token efficiency for complex explanations.
• HTML enables dynamic content like SVG and interactive widgets.

Method

Prompt LLMs to generate HTML artifacts for explanations, specifying desired interactive elements, styling, and content focus (e.g., "Render the actual diff with inline margin annotations, color-code findings by severity").

In practice

• Request HTML for code reviews to include inline annotations.
• Generate interactive security exploit explanations with HTML.
• Use HTML for technical documentation requiring diagrams.

Topics

• HTML Output Format
• Large Language Models
• Claude Code Team
• Markdown Efficiency
• Interactive Explanations

Best for: Machine Learning Engineer, NLP Engineer, Prompt Engineer, AI Engineer

Related on AIssential

• Using Claude Code: The unreasonable effectiveness of HTML — HTML offers richer, more readable, and interactive outputs for AI agents compared to Markdown.
• Claude Code Power Tips — Claude Code acts as an autonomous coding agent, transforming descriptive prompts into functional data science solutions.
• How to Convert Website to Markdown for LLM Training Data (2026) — Converting websites to Markdown is presented as the "gold standard" for preparing web content for LLM training, custom models, and RAG systems due to its ability to strip noise, preserve semantic stru…
• markdown-svg-renderer — The "markdown-svg-renderer" enhances markdown viewing by interactively rendering SVG code blocks alongside their source.
• How to Effectively Review Claude Code Output — The article highlights a critical shift in software engineering, where the bottleneck has moved from producing code to efficiently reviewing the extensive output generated by coding agents like Claude Code.
• Markdown is for Thinking. HTML is for Showing. — AI output needs a two-stage workflow, distinguishing between raw ideation and polished presentation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.