Tech • IA • Crypto
How We Claude Code
7/10TL;DR
Anthropic engineers are shifting toward HTML-based specifications and agent-driven verification to make AI coding workflows more reliable, scalable, and efficient.
KEY POINTS
Shift from Markdown to HTML specs
Engineers are increasingly replacing long Markdown requirement documents with HTML files that are denser and more interactive. HTML allows richer layouts, embedded visuals, and clearer structure, making it easier for both humans and AI agents to interpret complex product specifications.
Rising capability of AI agents
Advances in model performance are enabling agents to handle longer, more complex coding tasks. As agents run for extended periods, the risk of wasted compute and token usage grows, pushing teams to improve upfront specification and validation methods.
Interactive requirement extraction
Instead of rigid prompts, developers are encouraged to let models actively interview users to extract requirements. This approach reflects the idea that user intent is often implicit, and modern models can uncover it more effectively than static instructions.
“Bitter lesson” applied to prompting
Inspired by Richard Sutton’s “bitter lesson,” the approach emphasizes relying less on handcrafted constraints and more on model capability. Allowing models to generalize and infer requirements tends to outperform overly prescriptive prompt engineering.
HTML prototypes for rapid feedback
Teams generate multiple HTML design variants for applications, such as a bill-splitting app, enabling quick visual comparison. This makes it easier to provide feedback, especially when paired with screenshots and vision-capable models.
Agent-native verification systems
Verification is embedded directly into the application via DOM-exposed state and data contracts. Instead of traditional testing alone, agents can read structured data from the interface and validate behavior autonomously.
Three-layer verification workflow
The system supports verification across three modes: human-readable dashboards, agent-driven browser checks, and headless CI execution. This unified approach ensures consistency while reducing manual intervention.
Recording and sharing verification evidence
Verification runs can be recorded as video clips or artifacts, stored on platforms like S3, and shared across teams. This creates auditable proof of functionality and supports high-frequency deployment workflows.
Use of modern tooling
The workflow integrates tools such as Playwright MCP, Storybook fixtures, and structured schemas to define expected states and invariants. These elements allow agents to probe beyond “happy paths” and detect subtle failures.
Token efficiency through better specs
While HTML specs may initially consume more tokens, they reduce iteration cycles. Over time, richer specifications lead to fewer corrections, making the overall process more efficient.
CONCLUSION
The move toward HTML-based specifications and embedded, agent-driven verification reflects a broader shift in AI-assisted development, where systems are designed to collaborate with increasingly capable agents while minimizing ambiguity and manual oversight.