Three sports, one $15 rule

Every week I build a stack of things with an AI agent, and most of it dies in a closed terminal tab. This is the receipt — and fair warning, this was a loud week: 57 commits across 21 projects, including one game that turned into three.

The $15 game became a franchise

Last issue's quiet experiment — a basketball game where you draft a roster of all-time greats on a $15 budget — got rebuilt from the ground up. The 82-game season is gone; it's now a 16-0 gauntlet against four boss teams, with an era toggle, a deeper player pool, and chemistry bonuses for players who were actually teammates in real life.

My favorite exchange of the week: I asked, mid-session, "if the boss teams used their actual lineup numbers, would 16-0 even be possible?" The honest answer was no — so instead of inflating my side, the bosses' strengths are now derived from their real lineups and the difficulty is tuned openly in the math, not hidden in fudge factors.

Then it metastasized. A baseball edition (13-0 through October) was scaffolded straight from the basketball codebase, and for hockey I tried something new: I wrote a design brief before bed, let one model produce a complete design doc overnight, and had another implement it from the doc alone. I woke up to a playable 16-0 playoff gauntlet. Three sports, one roster rule.

A journal app my AI can actually read

I started building a personal iOS journaling app — native SwiftUI, with entries syncing through iCloud. The interesting decision was storage: I wanted the journal readable not just on my phone but by my AI assistant on the desktop. The answer was to put a small MCP server over the same database, so the assistant can query my entries the way it queries any other tool.

It's mid-build. Final development phase happens this weekend on a borrowed Mac Mini at my dad's place, which is its own kind of deployment story.

The boring part of running a video channel

My dive-footage shorts channel now has daily uploads scheduled out for weeks. The engineering this week wasn't editing — it was bookkeeping. With batches going up on a schedule, the real risk is posting the same clip twice, so the pipeline got an upload ledger as the single source of truth for what's been published, plus a playlist-based review flow so I stop opening twenty browser tabs to check clips one at a time.

Unglamorous, but this is the part that decides whether a channel survives unattended.

Reverse-engineering the cash-out button

New investigation: when you have a live bet and the game tightens, sportsbooks offer to buy you out at a price. That price comes from a function, and I want to know what it is. The week's work was a research plan — how to log offers systematically without doing anything that risks my accounts.

The practical wrinkle: I only bet on my phone, so the data source is screenshots. The plan is a collector that polls my photo library once a day and parses any new cash-out screenshots into the dataset. Nothing built yet — just 439 lines of plan — but the data collection design is the whole game here.

By the numbers

One big Wednesday spike — that's the gauntlet franchise day. An exam-prep tutor app also quietly racked up 11 commits and a 501-question bank this week without me touching a single file by hand.

Meta-confession: this newsletter's own pipeline sat uncommitted in a working tree for nine days. First thing this issue did was commit its own tooling.

Next week: the journal app comes home from the Mac Mini. — rickleberry