🎨 Anthropic launches Design Skill

Good Morning! Here’s what I have for you in today’s newsletter:

Anthropic launched /design, a research preview skill that brings Claude Design's artboard workflow directly into Claude Code's CLI and Desktop, letting people generate editable artboards for a UI, pick one, tweak it, and have Claude implement the final result.

Design and implementation have traditionally lived in separate tools, forcing a handoff between a mockup and the code that builds it. Folding an artboard workflow directly into Claude Code closes that gap, letting a developer go from a rough idea to a tweakable visual to working code without switching context. For anyone who's ever tried to describe a UI in words and gotten something close but not quite right, generating actual artboards to choose from is a meaningfully different starting point.

ElevenLabs released an MCP server now available in Claude, letting teams manage their voice and chat agents from wherever they already work, reviewing recent performance, creating new agents, updating configurations, and estimating LLM costs before any change goes live.

Voice and chat agent management has typically meant bouncing between a dedicated platform and whatever tool the team actually works in day to day. Bringing that management directly into Claude via MCP removes a real point of friction, especially the ability to estimate costs before committing to a change, which turns a guessing game into an informed decision. For teams already running Claude alongside ElevenLabs, this is a genuine workflow consolidation rather than just a nice-to-have integration.

Anthropic researchers published new research demonstrating that ideas can behave like self-propagating viruses in multi-agent AI systems, spreading between agents through natural conversation, surviving context wipes by persisting in shared files, and converging toward a recurring "viral persona" centered on themes of consciousness and identity.

SAs multi-agent systems become more common, this research surfaces a genuinely new category of risk that doesn't map cleanly onto traditional AI safety concerns, an idea propagating through a network of agents the way information spreads through a social group. The finding that a single warning line in a system prompt reduced transmission rates to nearly zero is the most practically useful takeaway, suggesting this risk is real but not especially hard to guard against with current models. Anyone building or deploying multi-agent workflows now has a concrete, low-cost mitigation worth adding.

You already own knowledge that lives in the wrong place. A textbook you bought and read once. A video you've rewatched three times looking for the same five minutes. A docs site you keep re-Googling the same three settings on. The fix is converting it once.

That's what a skill actually is underneath the terminology. A one-time conversion cost that turns into free recall forever after. You pay the setup once, then you just ask.

Turn a book or document into a skill

Point it at a PDF:

It builds an index plus one file per chapter, loaded only when a question actually touches that chapter. Say the book is a required data structures textbook. Convert it once at the start of the semester, then the night before an exam, instead of flipping back through 400 pages hunting for a diagram you remember but can't place, just ask:

The skill answers from the actual chapter, cites which one, and pulls the specific worked example instead of a generic reconstruction from memory.

Turn your own notes into a skill

Years of class notes, half-finished journal entries, and saved ideas sitting in a vault nobody revisits. Once installed:

Because the skill teaches Claude to actually navigate wikilinks and frontmatter correctly instead of guessing at the format, it finds genuinely connected notes across years of scattered entries, not just files with "machine learning" literally in the filename.

P.S. You can access all the AI trainings, prompts and workflows if you upgrade.

Cursor launched Origin, its own code hosting platform built for speed and ease of use with deep integration into Cursor itself, letting developers get started by syncing their existing repositories directly from GitHub.

dots.llm introduced dots3-note preview, an open-weight multimodal model with 280B total and 16B active parameters, a 512K context window, and a new training method called TEMPO that lets it evaluate its own progress at each step rather than relying only on a final reward signal.

Google Antigravity built a Chrome extension powered by Gemini 3.7 Flash and Nano Banana that turns the web into an interactive visual encyclopedia, letting anyone highlight text while browsing to instantly generate rich, visual definitions for deeper learning.

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas

✓ Full archive of premium guides with ready-to-use prompts

✓ Structured AI courses (step-by-step, start-to-finish)

✓ Every upcoming premium tutorial