Keep ultrathinking

Your Agent should use a File System This is a hill I will die on. Every agent can use a file system. The file system is an elegant way of representing state that your agent could read into context & allowing it to verify its work. 🧵on why and examples

we love launching the big stuff, but great products also rely on all of the small stuff. kudos to alex on the design team for design-plementing this (implementing and designing this at the same time! powered by claude code, ofc.)

I had some early access to Sonnet 4.5. It is a really good model. I saw especially big jumps in doing finance and statistics, which tend to get overlooked in the focus on coding.

Sonnet 4.5, it's incredible. Vibes on this model are off the charts. Try it out.

Along with Sonnet 4.5, all Claude.ai pro users now have access to a computer in Claude.ai that Claude can use to analyze data, make docs/sheets/slides, latex, react apps and more. vibe powerpoint and vibe excel are particularly awesome.

it's Claude sonnet 4.5 day!! to celebrate we made you a lil' demo with a bunch of cute touches. you can use it to watch claude generating UIs on the fly. 👏 to Martina Long Si Min Lee michael Sean Strong and many others for sprinting on this! claude.ai/imagine

This looks like an actually interesting and well thought out product designed to fit into people’s lives and emphasize their humanity — nice job!

We just enabled apt-get installs in Claude ‘s code execution feature! Have fun installing things :D

It would be cool to have some sort of a website where you and your friends could post photos and status updates to see what each other is up to

Keep thinking! Say no to slop.

New day new model

We’ve rolled out the ability for Claude to use Skills. Files are a wonderful way to encapsulate information that can be pulled into the context window as needed, as well as code to accomplish tasks. Skills are files the user can add to extend Claude in this way!

Claude + skills (canvas-design) made me this lovely coffee table book... think I'm gonna have to cmd+p this 🤓

Claude Skills is context engineering, not just a feature. The problem: Claude's default window is ~200K tokens . You can't dump all your docs/standards/workflows into prompts without degradation. Skills uses progressive disclosure - 3 tiers

we put claude code on your phone so you never have to stop vibeslopping to go to the toilet.

Slowly, and then all at once

Agents are becoming increasingly easier to build and make more powerful. I'm curious to learn where people think we can simplify Agent primitives and interaction patterns to give developers enough access to the important stuff without over-complicating the "low level" details

We never before needed to integrate every api on the fly, on behalf of and authenticated by a user. This is the MxN problem — how do all the different models integrate all the different APIs? This is what MCP solves.

Karina Nguyen you're absolutely right