I use Google Analytics to understand what gets read, shared and clicked here. It only loads if you accept — nothing third-party runs before that. How Google uses this data
Let’s learn in public — with the people and machines of this era.
A personal, independent commitment: I turn my experience with AI into language and give back what I find, as an ongoing book in situated chapters with agentic development as its open workshop. Each article is what I thought at one point of the journey, not a final word; the Library keeps the books that taught me. It goes on for as long as I can sustain the tools and infrastructure behind it — take whatever helps you.
Each one dated: what I thought, where I was
Large language models (LLMs) shortened the time between forming a question and receiving an answer. What I am learning now is about the interval after that: noticing what the answer changed in me before asking for another one.
What you'll walk away witha one-minute experiment for observing your attention between two answers, and a visual way to mark the environments where you think with AI.
▶ video in the article
Garry Tan breaks an AI-native company into four pieces that are all files, and a fifth one underneath: the brain. I took that breakdown and ran it against a loop I had already built without having seen the talk. Three of the four matched. The fourth and the brain break differently.
What you'll walk away withthe four pieces of his company are files, which is why you can rearrange them; the brain is what makes the rearrangement teach you anything — and the discipline holding both up is to never do one-off work.
▶ video in the article
Three agentic projects in parallel, on top of a full project calendar, and a bill of USD 4,207 covering two of them. The money is the least interesting number in it: what August actually cost was deciding the same thing over and over without noticing.
What you'll walk away withhow to notice you are burning hours and money on the wrong thing — three signals that need none of my tools — and what to do in the minute you notice.
▶ video in the article
I asked for a parody of a song I like, about my own work. Then I went image by image to check what each one actually referred to — and the checking is the article.
What you'll walk away withthat a prompt and its output prove nothing on their own; what makes the output worth publishing is going image by image and naming what each one refers to.
▶ video in the article
The badge for five years at AWS arrived six months late, and it made me count. This is not about the milestone — it is about the turn underneath it: what I went looking for in 2021, what I found, and why the work taking me further now is not the work I am assigned.
What you'll walk away withwhat repeated all three times the work stopped taking me forward: what I changed was the object of the work, not the company.
What got built, what got cut and why, what it costs a month, and the load-bearing decisions on record as ADRs. The whole setup, not only the conclusion it reached.
The first post. Why this space exists, and what I'm promising you — the honest version of a mid-career transition into AI Engineering.
What you'll walk away withwhat this space is, and the four promises behind it.
Open source for you to study, clone and use. It grows as automations graduate.
This site — a static React/Vite SPA, built in an agent-first loop with Claude Code.
A static SPA (no backend) served from S3 + CloudFront and provisioned with Terraform. Delivery is plan-first, with blocking CI gates and deploy on merge. The repo is the source of truth.
What you take awaywhere to start reading the repo, and what to reuse: the ADRs record the load-bearing decisions, from plan-first to deploy — with the trade-off each one cost.
The Claude Code plugin my agent loop took shape in: review personas that check an agent's work before anyone trusts it, and skills that document the conventions the agent follows.
Where the plugin is installed, a reviewer judges the MR in a fresh context and a hook refuses irreversible actions before the command runs, so the human keeps what is theirs. The tadeumendonca-skills plugin is disabled in this repository, and the method that runs here comes from the workstation at user level. The plugin is still published, but it is deprecated as a source, and its README points to tedeuxx/mhw.
What you take awaythe first packaged form of the loop, read through the decisions and trade-offs it records; the README opens by saying where the method lives now.