
AI Scaffolding Boils Down to Memory, Skills, Scripts
Thinking about simplifying what makes a functional AI scaffolding and how to think about building it up. Seems like you can reduce most approaches into three components: 1. A memory system: Contains relevant context about the domain the AI is operating on. What's going on in your business, your projects, your people. Most people use markdown files (CLAUDE.md, AGENTS.md). More complex setups exist depending on the use case. 2. Skills: a specific workflow or specific knowledge, encoded once and reused when the task calls for it. Drafting a newsletter, closing the books, processing call notes into next actions. 3. Bespoke software: where the work can be made fully deterministic, just have the AI write code. The model is great at judgment and synthesis, bad at anything that needs the same answer every time. If a script can handle it reliably, write the script.
Leverage Comes From User Scaffolding, Not Model Choice
Something I keep noticing: two people with the same AI model produce wildly different output. From my recent piece on why: "Two people with the same model and the same harness will produce wildly different results depending on the local knowledge...
Five Core Form Factors Define Next‑Gen Agentic AI
I spent some time going through Garry Tan's GBrain. I want to pull out what I see as the general form factors and what's interesting there as someone who is non-technical and doesn't work in VC. I think a lot of...
AI-Generated Album Impresses: Music Has Evolved
Haven't checked it out in a while but pretty amazed by where we are at with AI music. This is a pretty good album top to bottom https://t.co/dKg5ZGHBrf
Maximize AI Impact by Fixing Bottlenecks, Not Uptime
Eli Goldratt's book, The Goal, was famous for its (then unpopular argument) that keeping every machine running 24 hours a day, the metric most plant managers cared about, was actively making factories worse. I suspect we're seeing the same fallacy...
Opus 4.7 Reaches Usage Limits Much Faster than 4.6
Hitting my limits a lot faster in CC on Opus 4.7 than 4.6 I’m noticing. Was pretty rare that I’d every hit my 5 hour window limit but have hit it 3 times this week. Not doing different work and not...

Use Claude Code’s /Insights for Personalized Improvement Tips
If you've been using Claude Code for more than 2 weeks and haven't use the /insights command then I would highly recommend it. It produces a detailed report based on your specific usage with tips to improve. I always get 1...
Too Many AI Sessions Overwhelm Cognitive Capacity
For people running 10+ Claude Code or Codex sessions at the same time, curious what your experience is in terms of cognitive switching costs? I find that I can run like 2-5 sessions in parallel depending on the complexity. If I...
AI Mirrors Nuclear, Electricity, Press: Power, Infrastructure, Democratization
Just riffing on some AI metaphors/analogies floating around right now as a way of thinking through things. Here are some dominant ones: Nuclear weapons / Cold War arms race. AI as a dangerous new force that concentrates enormous power in the hands...
AI's Future Mirrors Cold War Socialist Planning Debate
Everyone loves the nuclear analogy for AI right now. It presses all the sexy buttons: power, destruction, governance, etc. Fair enough, but here's another Cold War analogy I've been noodling on: the socialist calculation debate. In the 1920s Mises argued socialism...

AI Mirrors Brunelleschi: Innovate On‑the‑Fly, Defy Limits
One historical analogy I keep coming back to in thinking about how AI is reshaping knowledge work is the Master Builder, most famously embodied by Brunelleschi In 1418, Florence announced a competition to finish its cathedral, but there was a wrinkle....

Layered CLAUDE.md Files Provide Contextual AI Guidance
I think the top tip I've picked up since I started using Claude Code: write a CLAUDE [dot] md at every level of your file structure and be intelligent about how you nest folders. Claude Code automatically loads the Claudemd file...

Nano Banana Cuts Concept Sketch Time From Hours to Minutes
As someone who spent years wishing I was better at drawing little conceptual images like these for articles or presentations, the ability to do this with Nano Banana now is absolutely amazing. This was maybe 5 minutes and 3 versions and...
AI-Generated Checklist Demystifies Taxes and Catches Errors
Wins from using Claude Code for my taxes for the firs time: First, the document upload checklist: A while back I uploaded my prior year's 1040 to Claude and asked it to help build me a tax prep checklist: every document I'd...

Treat AI Like a General Contractor, Not an Architect
My current favorite metaphor I've found for working with AI is the general contractor. Architect doesn't quite fit because it's too hands off. (This is maybe not the case for software where the models seem better) A GC is on site. They...