https://t.co/lAYTISRFRE #ClaudeCode

I opened 62 PRs with this over the weekend. 39 I merged, 21 I discarded. So it isn't perfect yet, and problems are both in 1) identifying real issues to fix 2) writing good code. But the orchestration part is starting to get solid.

It's not only about better models. Better models enable more reliability, which opens the door to parallelization. Git worktrees + atomic issue tracking + LLM and static quality gates = safe parallelization.

What if you could have 5 AI agents working on your code base at once: fixing bugs, simplifying code, and building features. All while you focus on coming up with new things to build? That's the promise of agent orchestration. Here's the loop I'm trying.

The more you know about a domain, the better you can guide the AI. https://t.co/g4Yi9dijbn

AI works best when you can verify the work. I can spot check the Amazon purchase history analysis. I can test if the server hardening actually worked and make sure I wasn't locked out along the way. It works even better when AI can verify its own work.

I've been using AI as my first stop for any problem I need to solve. This week: auditing my Amazon purchase history and hardening my home server. Here's what I learned about where AI actually adds value 🧵

Building for early readers means every interaction matters. We need to keep kids engaged, but challenged. https://t.co/fNtj7R50eo

The pedagogical goal: scaffold heavily at first, then progressively remove support as kids build confidence and motor memory.  Same letter, 4 different challenges. Kids can practice 'a' on rails in week 1, then draw it from memory when they've worked up to it over time.

Mode 4: Free draw No guide, no constraints. Pure recall from memory.  We even detect common mistakes (like mixing up b/d, p/q) so we can give targeted feedback when kids confuse them up.

Mode 3: Trace with guide The training wheels come off. Kids draw freely over a visible letter outline. No constraints, just a reference and their own motor control.  We verify it looks right using handwriting recognition.

Mode 1: Rails with guide Kids drag along the exact letter path with visual support. The path constrains their finger and they literally can't go wrong. Mode 2: Rails without guide Same constrained path, but now the letter is invisible. Motor memory kicks in.

We are shipping 4 drawing modes for letter practice in @ClassDojo Sparks.  Why? Because handwriting is crucial for reading development. It activates motor memory that helps kids internalize letter shapes.  Here's how we're making it progressively challenging:

Alright @explodeybook, you warned us that This Book May Explode, but you didn't warn me that my son would be shouting "by the power of corn dogs" all day long.

When we were first hiring engineers for this team, quite a few of them wouldn't talk to us after we told them we weren't going to use Unity to build it. We built a team of crazy good engineers and I'm excited they're showing off what web tech can do in 3D.

This has been an incredible project to work on. Data team built a platform where we can easily pull in new metrics from anywhere in the business. Each week, we can see exactly what's happening. This let's us scrub metrics with too much variation and focus on the important ones.

Class Dojo have launched a new platform- ‘Class Dojo Islands’. It’s an online virtual space where they can meet their friends and play. Year 4 made this video into show how it works:

Ed tech startup ClassDojo spent 11 years winning over teachers with its classroom app, today reaching 51 million kids. Now valued at $1.25B by Tencent, ClassDojo's unveiling its next act: an ambitious virtual world it claims will come parent-approved.

@ClassDojo, you da real MVP 👊🏼 What other EdTech company provides the goods on a regular basis like ClassDojo?? Not only is their content high-quality & research-based, but it's all packaged beautifully, has its own aesthetic/branding, and is actually kid-friendly! Dojo FTW!

I wrote about how we prioritize bugs at @ClassDojo, how it avoids wasted effort, and how it keeps our quality bar high