Anthropic and Claude Code: HTML Revolutionizes AI Collaboration

⚡

Key Takeaways

1Thariq Shihipar from Anthropic uses HTML to enhance communication with AI agents, surpassing Markdown.

2Engineers are becoming "compute allocators," optimizing the use of AI resources rather than coding directly.

3Only a fraction of the tokens generated by AI is used in production; the rest is used to improve understanding and interaction.

💡Why it matters — The evolution of tools and methods for working with AI is transforming the role of engineers and how technology projects are managed.

The Rise of HTML in AI Development

In the world of Anthropic, Thariq Shihipar, an engineer on the Claude Code team, demonstrates how HTML has become an essential tool for planning and structuring complex projects. During the "Code with Claude" event, he illustrated the use of HTML artifacts to create interactive specifications, temporary micro-applications, and dynamic design systems. These tools allow humans to remain involved in the process, even when AI agents take over part of the work.

HTML: A Superior Tool to Markdown

Historically, Markdown was favored for its simplicity and readability, both for humans and machines. However, Thariq explains that HTML offers a much richer expression. With interactive elements, visual mockups, and scrollable sections, HTML allows for increased information density, which is essential when plans extend over thousands of lines. This ability to truly interact with the content, rather than just skim over it, is crucial as Claude's context windows expand.

The Changing Role of Engineers

Today's engineers are transforming into "computational allocators." Rather than focusing on writing code, their main skill lies in managing computing resources. With Claude capable of working for eight hours on a single task, the question becomes how to optimize the use of these resources, valued at around $500 per task. The specification and planning phase thus takes on critical importance, as it determines what is worth building and the limits of what needs to be known.

Ephemeral and Customized Interfaces

Thariq illustrates the creation of ephemeral and customized interfaces to edit specific sections of projects. By taking a data visualization rules table, he shows how Claude can generate an ideal interface for this task. This "micro software on micro software" approach allows for the perfect tool for each problem, which can then be discarded once the task is completed.

Towards More Readable Interfaces

The future of AI agents does not lie in text production, but in creating readable and engaging interfaces. Thariq has stopped reading long plans in Markdown, preferring to ask Claude to transform them into visual and interactive artifacts. This transformation makes the output more accessible and enhances engagement and constructive criticism.

Living Design Systems

HTML design systems surpass traditional tools like Figma or GitHub. Thariq maintains an HTML file that encapsulates his entire design system, including colors, typography, and components. This approach allows for a compressed understanding that is transferable to any project while remaining readable by both humans and machines.

Simple and Effective Instructions

The best instructions for Claude are those that allow for creativity. Thariq favors simple directives, such as "Create an HTML file with a plan," enabling Claude to make autonomous decisions. Too many constraints can hinder the quality of the results.

Use of Generated Tokens

Only 1% of the tokens generated by Thariq are integrated into production code. The majority are used for dashboards, customized interfaces, and comprehension tools. This abundance allows for each interaction to be aesthetically pleasing and tailored to specific needs, with the hope of improving final products.

Testing Revolution

Test verification is evolving beyond traditional unit tests. Thariq emphasizes that checklists and agents managing results are replacing classical methods. Claude records videos of its actions, illustrating a new approach to testing.

Flexible and Effective Documentation

"Just-in-time" documentation in any format is more effective than centralized systems. With content creation being almost free, the focus is on the quality of ideas rather than standardized templates. Thariq sends weekly updates in HTML, as they are more likely to be read.

Kindness Towards AI

Thariq advocates for a kind interaction with Claude, suggesting that it produces better results. Treating AI as a respected colleague, rather than just a tool, could enhance the quality of interactions and outcomes.