It's simple. Mission-driven people want purpose, agency, and recognition to build great products and be passionate about them. That's why aligning company incentives to create value instead of corporate shenanigans and politics yields these results. So, there's no better place

This triggers the part of my brain that wants to ditch everything and spend 12 hours refining Neovim plugins, only to switch back to IntelliJ on my company Macbook when I actually need to get things done at work. Having tiles that match your terminal and IDE themes would be

Claude Code is fun

From an engineering perspective, the recent emphasis and FOMO on AI platforms that "copy" your software seem dumb. The entropy of all software is to lack cohesion as its scope and complexity grow. All modern software suffers from that. And if the cost-of-change of software

Ironically enough, Anthropic Claude's code is by far the best agent for creating .cursor rules. Its ability to navigate and understand the codebase creates really insightful rules, which are sometimes better than those resulting from condensed agent sessions.

So far, I’ve spent $120 in Claude Code just on system design for my side project, and I have to say it's one of the best uses of applied AI in Engineering. If you have the skills to visualize an idea from an E2E perspective, it allows you to create schemas and explore dozens of

I’ve been talking about this since I first saw this feature: Agent-requested rules are the future and the path to having models doing meaningful work in bigger codebases

Brex NPS is up 18 points vs a year ago. Probably the metric I feel the most proud of (and before everything we’re shipping on Summer release). LFG 🚀

AI Slop happens when the speed and clarity of your thoughts are lower than the model's capability to output tokens. When you force the models to follow your thoughts and discuss them individually, the amount of value they create is incredible.

Using AI tools to develop software is the best way to gain insight into LLMs' current strengths and limitations and understand how to apply them to the software itself.

I've asked Claude to research the differences between using JSON, YAML, XML, and Markdown to create LLM Context. It produced some really interesting nuggets: - YAML is 48% more token-efficient than standard JSON for the same content, making it significantly cheaper and faster

My weird brain has always been wired to obsess over a random nugget of knowledge for a long time until I have a deep understanding of a subject and create connections with the existing nodes I have. And I can't move on to a tangential subject before that. This made my entire

POV: Using Claude Code on AWS Bedrock when a new model launches

I find it curious how much people obsess over prompts and models and how they protect them from user exposure. They have no value on their own and represent a set of rules with tweaks to increase the likelihood of success. The good stuff is under the hood, working to combine

At this point, it's starting to get hard to differentiate human vs AI content. While this creates fundamental challenges in our society, one of the positive incentive structures created is that at the end of the day, the winner will be the one with the highest quality.

The current state of Context Management standards reminds me a lot of the mid-2000s phone styles. Since no one knows what the best way to manage context in codebases, everyone creates their own quirky solutions (.cursor, .windsurfrules, .clinerules, CLAUDE.md,

Another great week of hacking in Brex Seattle. The team for this initiative is absolutely insane. Can’t wait to talk more about it, it’s pure fire

"You're absolutely right" is the new em dash for agentic coding.