I manage developers for a living. In late 2025, every leadership meeting I sat in had the same agenda item: "AI-assisted development." Not should we adopt it. That ship had sailed. The question was how, and nobody in the room had a good answer because nobody in the room had actually used the thing.
Including me.
I had two options. Option A: read some blog posts, form an opinion from a safe distance, and tell my teams what to do with the confidence of a man who has never been wrong about a thing he hasn't tried. Option B: actually use it. Every day. On everything. Until I had something worth saying.
I picked B. Over three months I built 40+ projects: made a website specifically designed to annoy my friend Tim, deployed five Lambda functions so I wouldn't have to walk five feet to my fan, committed API credentials to a git repo because it was easier, wrote a document called SECOND_AUTH_INVESTIGATION.md, and accidentally built an AI music platform when I asked for a drum loop. None of these are finished. All of them taught me something.
This site is the journal.
The Ladder I Didn't Know I Was Climbing
Around project 15, I noticed my relationship with Claude Code had changed. I wasn't typing the same way. I'd gone from asking permission to giving orders to setting up entire projects and checking in later. It mapped to something:
| Level | What It Looks Like |
|---|---|
| 1: Prompting | "Can we use CDK?" (asked like I need permission) |
| 2: In the Loop | Reviewing every output, fixing every mistake manually |
| 3: On the Loop | "Build the auth system," reviewing the PR, not the keystrokes |
| 4: Managing the Loop | Writing the CLAUDE.md, the settings, the conventions. The setup IS the work. |
| 5: Orchestrating Loops | "Do the corolla next," five words, complete car research report |
Each post on this site is tagged with a level. The companies that stay at Level 1, treating AI like a fancy autocomplete, leave 90% of the value on the table.
The Pattern
Every project followed the same arc, which I found comforting and then depressing:
Ambitious idea. The ambition scales with confidence. Early on it's "make a grocery list." By month two it's "reverse-engineer a proprietary Bluetooth protocol and build a serverless IoT cloud bridge."
Surprisingly fast progress. Claude gets you 70-80% of the way shockingly fast. The first time this happens you think you're a genius. By the tenth time you realize the tool is a genius and you're just the person who typed "build me a thing."
Vibe code into the weeds. "Yeah just add that." "Can you fix that real quick." Suddenly you have a semi-functional thing held together with optimism and // TODO: fix this later. This is the phase where your engineering instincts should be screaming but you're having too much fun to listen.
Move on. Because the point was never the product.
The ratio of "me typing code" to "me reviewing code" shifted steadily toward review. By month three, my contribution to most projects was questions, judgment, and the occasional all-caps NEVER DO THIS in a CLAUDE.md file. Claude's contribution was everything else.
The posts are sorted by level. Start anywhere. But if you start at 1, you'll watch someone go from asking a coding tool for permission to orchestrating eight AI agents in parallel without writing a line of code. That arc is the point.