What happened

Anthropic has released an early preview of the /design command in Claude Code, according to The Decoder and the owned-news signal. Developers can generate UI mockups as artboards directly in the terminal before writing any code.

The Decoder reports that Claude reads the existing codebase to match current UI styles, enabling a visual planning phase before code implementation. Anthropic is embedding a visual design workflow into the terminal-based coding agent rather than requiring a separate design tool.

Evidence

• Anthropic released an early preview of /design in Claude Code for terminal UI mockups.

  The Decoder · attributed

  Anthropic has released an early preview of the /design command in Claude Code. Developers can generate UI mockups as artboards right in the terminal before writing any code.

• The /design command reads the codebase to align mockups with existing UI styles.

  The Decoder · attributed

  Anthropic introduced a /design command in Claude Code that allows developers to generate UI mockups as artboards within the terminal. The tool reads the existing codebase to match current UI styles, enabling a visual planning phase before code implementation.

• Claude Code cost efficiency may depend more on effort-level settings than on model selection.

  Towards AI · attributed

  Claude Code cost efficiency is driven more by effort-level configuration than model selection, with Sonnet 5 becoming the default agentic standard in late 2026.

• DeepSeek Harness is a model-agnostic agent runtime that has gained traction on GitHub.

  Towards AI · attributed

  DeepSeek released 'Harness,' an MIT-licensed, model-agnostic agent runtime built on the Cordis architecture, which has rapidly gained traction with over 33,000 GitHub stars.

• Major providers are testing different compute paradigms for giving agents computer access.

  Ethan Mollick · attributed

  The industry is currently testing three distinct architectural paradigms for AI agent persistence: local execution, ephemeral cloud sandboxes, and persistent web-based instances.

• Anthropic's text watermark indicates likelihood, not certainty, and is sparse in code and factual text.

  Techmeme · attributed

  Anthropic has detailed the technical specifications of its text watermarking system, clarifying that it indicates likelihood rather than certainty of Claude's involvement. The system is intentionally sparse in code and factual text to minimize interference wi

• Anthropic is launching a watermark detection API for third parties based on a modified SynthID approach.

  The Decoder · attributed

  Anthropic is launching a watermark detection API that allows third parties to verify if text was generated by Claude. The system leverages a modified version of Google's SynthID method, adjusting token selection randomness to embed detectable signatures witho

• Claude Code prompt caching is invalidated by switching models or altering early context.

  Hacker News AI · attributed

  Prompt caching is fragile to structural changes, meaning switching models or altering early context invalidates the cache and incurs full prefill costs.

Why it matters

Teams using Claude Code may reduce tool-switching between design and implementation if mockups stay aligned with existing codebase styles before any code is written.

Limits and uncertainties

The /design command is an early preview; production stability and full capability scope are not established in the packet.

The Towards AI MCP server configuration article is listed but carries no analysis or excerpt detail in the packet.

Anthropic's watermark signals likelihood rather than certainty and is sparse in code and factual text, limiting provenance use for technical content.

Practical implications

Treat effort-level settings as a primary cost control variable when running long Claude Code sessions, not only model tier selection.

Avoid switching models or rewriting early session context once a Claude Code session starts to preserve prompt-cache efficiency.

Evaluate agent harness design separately from underlying model choice when comparing Claude Code against alternatives such as DeepSeek Harness.

What to watch

Whether /design moves beyond early preview and how mockup artboards hand off to code generation inside Claude Code.

How Anthropic's watermark detection API performs in moderation pipelines given sparse signaling in code and factual text.

Which compute persistence model teams adopt for agent workflows as local, ephemeral, and persistent paradigms diverge.