Ref: https://job-boards.greenhouse.io/anthropic/jobs/5448167008
Why Anthropic
I’ve been using Claude daily since March 2024, when Claude 3 came out, and I’ve used it across my whole life: as a husband and dad to a 9-month-old (and two cats), and as CEO of DeepSource and a prolific designer/developer. I’ve been consistently impressed by both the models and the ingenuity of the human-agent interfaces that have come out of Anthropic.
Using Claude this long, I’ve come to believe that Anthropic is the only lab with a clear, coherent vision of what a post-powerful AI (to borrow Dario’s phrase) world looks like for both businesses and consumers, and a roadmap for how to get there, with new product experiences that keep up with the model leaps while taking safety equally seriously. Shigeru Miyamoto of Nintendo said a good idea solves multiple problems at once, and I’ve seen several ideas like that come out of Anthropic in the last couple of years, big and small.
Two of my favorite examples are skills and background watchers in Claude. The combination clicked for me when I was building personal apps and was able to create novel experiences that wouldn’t have been possible before (happy to show a walkthrough in person). I’m also a big Claude Code fan, and I’m consistently impressed by the UX patterns invented in the product that defined agentic coding for the industry (big ones like agent teams, subagents, auto mode; small ones like /btw and auto-compaction). I find that inspiring.
Since I sold DeepSource, I’ve been thinking about a few things: how AI changes the economics of software (I wrote my thesis in this essay: https://sanketsaurav.com/writing/end-of-rationing), how we steer toward a world where everyone has access to powerful AI and uses it for real human progress, and, tactically, how to do this while making sure the people building it most responsibly stay at the frontier. Several of my fellow YC founders now work at Anthropic, and talking to them, it’s been clear to me that my thinking on these aligns with Anthropic’s the most.
Models are getting better much faster than the tools we have to use them. Working at Anthropic, I’d like to close that gap: go deep on human-agent interaction and invent new patterns and primitives so we can build new and better products that let our users get the most out of Claude’s capabilities. Products are the biggest lever for getting powerful AI into everyone’s hands. That’s why I’m excited about this role.
Additional Information
I’ve used Claude Desktop for a few hours a day, consistently, over the last year, and I’ve used it to drive a substantial chunk of my work as CEO of DeepSource, including running the acquisition process end-to-end by myself in a single Claude thread that was open for 5 months.
From that first-hand experience, I believe the models are improving faster than the tools we use them through. Here are three ideas for closing that gap.
Claude that compounds for the user, not just for Claude
- Today, Claude’s persistent layers, like memory and project context, help Claude learn about the user and get progressively better. There’s nothing that helps me, as the user, get progressively better as I use Claude. Claude’s understanding of me compounds, but my own understanding doesn’t.
- I built a workaround for this with Obsidian: I’d write freeform in my daily notes folder about outreach ideas, offline research, and general strategy. Then I created a “distill” skill in Claude that ran on my daily notes and added them to a “knowledge base” folder in a structured, backlinked format optimized for me to read. Claude wasn’t allowed to touch my daily notes, and I wouldn’t edit the knowledge base myself. This let me think independently, while Claude augmented my thinking and understanding when I read the knowledge base every other day. I could refer back to a page Claude had written or share it with others.
- A knowledge base/second-brain layer built into Claude, similar to my workaround, would solve this.
Human-agent collaboration
- The primary way users work with Claude is through the chat box, in a turn-by-turn conversation. This works for most use cases, including coding, but a more sophisticated human-agent interface could amplify what’s possible with Claude.
- For example, when users create documents or other artifacts with Claude, they can comment inline to ask questions or request changes. That’s great, and a step beyond copy-pasting into the main chat box. To take this further, Claude could proactively “watch” the changes users make to an artifact and “suggest” what to do next, or what to change elsewhere to keep things coherent. Pulling on this thread leads to product experiences where users can use Claude far more effectively to do their jobs.
Connectors and webhooks in Projects
- I’ve found Claude Projects a great abstraction for dividing and keeping track of the different kinds of work I do with Claude. Most of my projects are long-running and accrue a lot of context and memory over time. For instance, I have projects called “DeepSource Ads,” “DeepSource Sales,” and “DeepSource PM” for company work, and “Body Recomposition” and “Personal Finance” for personal work.
- Almost all of the work I do in these projects relies on live data from external tools. Individual chats can use connectors, but a project’s context can only hold documents from a few sources (files, text, GitHub, Figma, Drive). I’d love to attach connectors to the project itself, so every chat in it starts with live access to the data the project depends on.
- Taking this one step further: projects can already run scheduled tasks. Triggering chats from external webhooks would make them event-driven, not just time-driven, and much more valuable.