The Collective Catch-Up
Anthropic launches Claude Design — visual work, in the conversation.
A new Anthropic Labs product lets you collaborate with Claude on designs, prototypes, slides, and one-pagers without leaving the chat.
On April 17, Anthropic introduced Claude Design — a research-preview workspace for creating polished visual work directly with Claude. It's powered by Opus 4.7's vision model and is rolling out gradually to Pro, Max, Team, and Enterprise subscribers.
The pitch is simple: instead of context-switching between a chat window and a design tool, you describe what you want and iterate on a live canvas. Worth watching for anyone shipping their own marketing pages, decks, or product mockups.
Read the announcement: anthropic.com/news/claude-design-anthropic-labs. If you get into the preview, share what you build in #03-show-and-tell.
Opus 4.7 lands — built for long-horizon, autonomous work.
On April 16, Anthropic released Claude Opus 4.7 across all Claude products, the API, Bedrock, and Vertex. The headline gain is on the hardest software-engineering tasks — the kind that previously needed close human supervision.
Anthropic also flagged improvements in instruction following, multimodal/vision handling, and self-verification before reporting back. Independent coverage (Help Net Security, MarkTechPost) emphasizes the same point: this is a model tuned for longer, less-supervised runs.
Heads up: Sonnet 4 and Opus 4 (the original 4.0 models) are deprecated and retire on June 15, 2026. If you're pinning a model in production, plan the migration now.
Anthropic locks in gigawatts of compute — with Amazon, Google, and Broadcom.
Two infrastructure deals in April reshape the supply curve behind Claude — and signal where the company is betting capacity will need to be over the next few years.
On April 20, Anthropic announced an expanded agreement with Amazon to secure up to 5 gigawatts of capacity for training and deploying Claude — including new Trainium2 capacity coming online in the first half of 2026 and nearly 1GW total of Trainium2 and Trainium3 by year-end.
Two weeks earlier, on April 6, Anthropic signed with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, expected to come online starting in 2027.
Sources: anthropic.com/news/anthropic-amazon-compute · anthropic.com/news/google-broadcom-partnership-compute.
Only
Drop
Claude Code Skills are the highest-leverage feature most people aren't using.
Skills extend what Claude can do: write a SKILL.md file with instructions, and Claude loads it into its toolkit on demand. Anthropic introduced Skills across Claude in late 2025 and has expanded them through early 2026 — and this month the community coverage caught up.
Multiple guides published this week (DEV, agentincome.io, LaoZhang AI, PopularAiTools) all land on the same advice: start with the bundled official skills before chasing third-party marketplaces. The two most-recommended starters right now are webapp-testing for frontend QA and document-skills when PDFs or decks are part of the job.
If you keep pasting the same prompt or stuffing more into CLAUDE.md, that workflow probably wants to be a Skill. Anthropic's docs are at docs.anthropic.com/en/docs/claude-code/skills.
Computer use moves into the desktop, on the user's machine.
Coverage from creati.ai on April 4 framed it well: AI is "rapidly pivoting from passive content generation to active task execution," and Claude's expanded desktop control is one of the clearest signals. The frontier of agentic AI now includes Claude reading your screen and clicking the buttons.
Useful patterns to think about: pulling reports out of vendor dashboards that don't expose APIs, running through onboarding flows to test sign-ups, and handling repetitive admin tasks across web apps that nobody loves.
The interesting question isn't "can it click the button?" anymore — it's "what's the smallest reliable loop you'd hand off, and how do you supervise it?"
Two practical cautions if you're trying it: scope sessions to disposable browser profiles, and put a human-in-the-loop checkpoint on anything touching money or identity. "Claude proposes, human approves" is a fine starting posture until you've watched the same routine succeed ten times in a row.
If you've shipped a computer-use workflow that's actually saving you time, post it in #03-show-and-tell — we may feature member experiments in a future Catch-Up.