Thoughts on product management, WordPress, WooCommerce, leadership, and coaching from Chris Lema.
I asked Claude to interview me and got twelve questions back. I answered two. What worked better was having it guess first and let me correct it. Here are the project instructions I use now, and the four files that come out of an interview.
In 1986, an information scientist found a hidden medical connection nobody had made — not by running an experiment, but by linking two piles of research that never cited each other. I set out to build that same instinct into our knowledge graph. Then I got told I was 60 years late. Then 12. Here's what that taught me about why your second brain shouldn't just sit there.
Carlos Perez's essay is the clearest map I've seen of where AI agent architecture is heading. Here's a walkthrough of it from inside a real production system — five agents on Mastra, seven days in — including the parts I got wrong along the way.
Most AI adoption stalls on a false choice: keep AI on a leash, checking every step, or let it run and hope. There's a third path — encoding your own judgment into the system so AI can run inside guardrails you built. Using a real example (surfacing partnership leads from LinkedIn), this post breaks down what it actually takes to turn instinct into a system.
Every service business is staring at AI knowing everything is changing and nothing is staying the same. That forces a bet: go premium, go volume, or shift to equipping. But all three bets share one dependency, and the same AI that made your expertise cheap is exactly how you feed it.
I sat down to learn Mastra with my Claude Code habits and my Cloudflare opinions in my back pocket. A few agents and a workflow later, I'd accidentally built a full coding harness. Here are the seven rules I refused to bend — and what each one looks like in running, open-source code.
I spent a weekend building a sales executive who doesn't exist, handed her fifteen secrets, and let an AI try to pull them out. The fake expert was the point: it's the first time an interview had an answer key, so it could earn a score instead of a compliment. Here's what testing an AI honestly actually costs.
I sat through a multi-hour AI training with 3,000 people and discovered less than 10% had ever tried a second tool. Comfortability keeps you stuck in the tool you started with. Exposure is how you get out, and here are two easy ways to choose it.
Ask an AI to build the same thing twice and you get two different programs, both working. So you can't test the code against one right answer. Here's how to test what it actually does instead, in five layers, with a Stripe webhook as the example.
It's not the model, it's the harness. Here's the one I built for content, plus a 7-minute demo of it running.
The real threat AI poses to your voice isn't reduced typing, it's the erosion of your taste every time you approve a suggestion instead of overruling it. Keep your voice upstream (the deciding) and confine AI downstream (the executing) by making four non-delegable calls before you generate a word.
You installed a Claude skill someone else published and it just sits there, never running. That isn't a broken skill, it's how skills load. Here's why it happens, how to tell, and two ways to fix it, including one that needs no editing at all.
You shipped something real and nobody knows it exists. Video used to mean gear, a crew, and an edit. Now it's a short script and about five minutes, which moves the hard part to whether you've got something worth watching.
I ran a product team that captured every customer question into a playbook before handing it off. That old loop is exactly how you make a cheap AI model carry expensive work.
My take on loop engineering: push judgment onto dumber, cheaper models and smarter code over time, so a long-running agent gets cheaper and more trustworthy with use.
I put AI to work on the boring part of my site: reading 74 posts to surface the topics, audiences, and cross-links that were hiding in them all along.
I open the hood on the Claude Skill that builds mirror assessments — the eight-step pipeline, the single spec it compiles, and the checker that won't ship broken.
I gave a one-page brief to a new Claude Skill and got back a working lead-gen quiz — a design doc and a hostable assessment. Here's how the skill does it.
Everyone needs a custom assessment, and when you think of your own experiences, you'll agree with me.
Five paradigm shifts from three years building with AI: design time over runtime, judgment over prompts, real feedback loops, conversation, and knowledge that's built, not stored.
A pizza shop's flyer contest shows why your AI scoreboard can climb while the work gets worse - and why a measure that can't be gamed is the real job.
Why storage is the wrong question for knowledge, and how decomposing content into a graph builds answers no document ever held. The MCode Coach story.
One prompt gets you something. Ten conversations get you something better. Here are the ten you should have with Claude before any code gets written.
Companies are blowing through AI budgets, but pricing isn't the real cause. Two habits keep my bills flat without rolling back what AI lets me do.
When you build with AI doing the content generation, the obvious move is QA at the end. But if that's your whole strategy, you sat down at the piano without picking a key. Here's how I'm thinking about evals for a new product — and why guardrails belong upstream, not downstream.