Thinking about creating a /pr skill. You invoke it to create a PR that: - Makes human review as fast and as simple as possible - Takes the best parts of /show-me by @dexhorthy - Uses your domain language (from /grill-with-docs) - Shows evidence that the change works (images, test output etc) - Assesses merge risk (is it a one-way door, or a two-way door?) IMO every model/harness I've seen creates garbage PR bodies - needs a skill here. WDYT?

So last week, I interviewed with Anthropic. The first stage of their interview process was a nonsense Leet code problem, with an interviewer that didn’t ask me any questions, and I’m positive he didn’t know anything about me. Later that day, an Anthropic researcher quit and said he was worried about the company causing human extinction. 🤣 Hard pass. Haha. I don’t think they liked me anyway.

Read the full AI SDK 7 announcement: vercel.com/blog/ai-sdk-7

My thoughts on eve vs AI SDK AI SDK is lower level foundation, focused on provider abstraction and related functionality, and as minimal dependency as possible. It is UI framework agnostic and gives you the full control, including provider-specific configuration options. The tradeoff is that you need to use other libraries (e.g. workflow SDK, chat SDK), wire everything together, define the deployment, think about the prompts, etc when building agents with AI SDK. This is where eve comes in - it takes care of wiring, integrations, basic prompt assembly and makes it super easy to create agents. Very excited that we now have this higher level framework with eve - a Next.js for agents!

It's awesome that people are actively asking when ai-retry will support AI SDK v7 Here we go: `npm install ai-retry@beta`

painful mistake adopting @EffectTS_ that we've finally unpicked always start with a managed runtime and build up from there avoids odd runtime bugs, layer duplication, easier flushing and simpler interop with vanilla ts code @SandroMaglione was right typeonce.dev/course/effect-b…

I’ve had immense difficulty refactoring with agents. The old code usage patterns are like a siren song for the agent. If you don’t check, you refactoring task MIGHT end up; Instead of A => B Doing A => A+B That is: agents think nothing of keeping the legacy code around for all sorts of reasons: fallbacks, shims, bridges, legacy… Even when you’re very explicit I’ve been through this cycle seven times in one project and I just could not get rid of the legacy code My final solution was a “delete first” approach If you delete the legacy code first and ask it to re-implement the same contract / passing tests with the new code, it can’t rely on the old code. It’s the only way I’ve been able to get it to reliably removal references to the old code.

Here's a deep dive into why "𝚞𝚜𝚎 𝚠𝚘𝚛𝚔𝚏𝚕𝚘𝚠" uses directives and why it's the best choice for the framework useworkflow.dev/docs/how-it-…

In 2025, the AI SDK went from 400k to 3.5 MILLION weekly downloads. And I've built a self-paced, interactive, 57-lesson course to help you master it. And for a few days, it's $99. 2025 is the year of TypeScript + AI. Let's go.