Anthropic quietly shipped a /design skill into Claude Code. No install. Update Claude Code, type /design, it's there. https://code.claude.com/docs/en/skills What it does: it pulls Claude Design's artboard workflow into Claude Code. You describe what you want. You get back a link to an editable canvas. A full landing page with dark and light mode, fonts, copy, layout, off one prompt. Then you can actually edit it. Accent colours, spacing, font scale, straight from the page. There's a sidebar on the right that'll feel familiar if you've used Figma. It's a Claude artifact, so you share it like a Google Doc. Whoever you send it to can leave comments on the canvas. I haven't put it through a real client build yet, so I'm not going to tell you it replaces your design tool. What I will tell you is the reason it's worth your time over Claude Design on the web. Two things. Both are about where it runs. 1. It has your context Claude Code already knows your repo, your notes, your brand files. One short prompt came back with the right copy plus the right numbers, because it didn't have to be told any of it. On the web you're starting from a blank room every time. 2. It can use your own skills Open the skills dropdown in Claude Design and you get Anthropic's defaults and nothing else. You can't add your own. In Claude Code every skill you've built is already sitting there next to it, so your own generators become part of the design step. And when you're done you just say deploy it. Claude Design hands that back to you to do yourself. So the rule of thumb: if the thing you're designing needs to know anything about your business, do it where your context lives. What would you point it at first, a landing page or something else?

It gives Claude Code a memory that survives closing the terminal. Free, Apache licensed, 90,000 stars. Works in Claude Code, OpenCode, Antigravity and OpenClaw. Here is the problem it solves. Every session starts from nothing. You explain the project, Claude works it out with you over an hour, you close the window, and all of that understanding is gone. Next morning you explain it again. Most people just accept that as how the tool works. claude-mem sits on five lifecycle hooks and watches the session while it happens. Tool calls get captured, compressed into short observations and written to a local SQLite database. Next time you open that project, the relevant history is already in context before you type anything. Setup is one command. Run npx claude-mem install, then restart Claude Code. Or add it as a plugin with /plugin marketplace add thedotmack/claude-mem followed by /plugin install claude-mem. The part that makes it work is how it reads back, and this is the bit worth stealing even if you never install it. It does not dump your history into context. It goes in three layers. Search returns a compact index of matching observations at roughly 50 to 100 tokens each. Timeline shows what was happening around one of them. Only then does it fetch full detail for the handful of IDs that actually matter, at about 500 to 1,000 tokens each. Filtering before fetching is where the saving comes from. My session this morning read 25,674 tokens of memory to stand on 204,350 tokens of previous work. That is 87% less than loading the lot. There is a privacy control worth knowing about. Anything you wrap in private tags never gets stored, which matters if your repo has client names or keys in it. Repo: https://github.com/thedotmack/claude-mem Docs: https://docs.claude-mem.ai The takeaway is not really the tool though. It is that context is a budget, not a bucket. Search first, fetch second, then you get the benefit of a long history without paying for it on every single prompt.