This week's video

Over the last three videos I covered how OpenClaw works, how agent memory systems are built, and how to give agents the ability to search and recall past conversations. This week is where it all comes together.

I build a personal meeting prep assistant — the same patterns behind my own AI assistant, Jarvis — using Mastra and TypeScript. A call gets booked, the agent researches who it's with, posts a prep brief to Slack, and schedules the follow-up. Tools, memory, webhooks, and scheduling — connected into something you can actually use.

The full source code is on my GitHub with a README that walks through setup. Watch it here →

What I'm watching

Google releases CLI for Google Workspace

First-party Google integrations for AI agents just got a lot easier. Google shipped a CLI for Google Workspace this week that makes it straightforward to connect agents to Calendar, Gmail, Drive, and Docs without wrestling with OAuth flows from scratch.

For anyone building a personal assistant like Jarvis, this is a meaningful unlock — your agent can read your calendar, search Drive, and interact with Gmail natively. I'm planning to wire it into Jarvis soon. Check out gws →

Qwen 3.5 Small Model Series

Alibaba released the Qwen 3.5 small model series this week — 0.8B, 2B, 4B, and 9B models built to run on modest hardware without API costs or rate limits.

This matters more than it might seem. People are already running Qwen 3.5 locally to power 24/7 coding agents — researching, building, and iterating around the clock with no API bills and no rate limits cutting long tasks short. The smaller models in this series bring that within reach for most modern machines, not just setups with a lot of unified memory.

If you want to wire local models into your own agent setup, Mastra has built-in support for it — all you need is an OpenAI-compatible local inference server.

The Claude-Native Law Firm

Zack Shapiro runs a two-person law firm that handles workloads a larger practice would staff 10+ people for. His writeup on how he actually does it is worth your time.

The short version: at 7 PM the night before a deal was closing, opposing counsel sent a last-minute demand letter. Zack fed the full agreement and letter to Claude, and it found that two of the buyer's proposed changes directly contradicted representations they had already confirmed in their own disclosure schedules. The deal closed the next morning on his client's terms.

But the more interesting part isn't the war story — it's how he built the system. He created custom Claude skills that encode his decade of legal judgment: how to weigh risk, what to flag, what format to use, what tone to take with clients. As he puts it: "The difference between a firm playbook and an individual lawyer's encoded judgment is the difference between giving someone a recipe and teaching them how to cook."

If you're building with Claude and you haven't thought about encoding your own judgment into a skill — not a generic one, yours — this is worth the read. The Claude-Native Law Firm →

If you are a team looking to adopt AI like Zack, I can help. Find out more on my website.