What I'm building with AI in the loop, mapped and updated weekly.

THE STATIONS

Three radio stations with nobody on the payroll.

One of them is a thank-you gift. My trio wanted to give the local startup incubator something to play in the office, so the office got its own station, with its own announcers, news desk and serials. ScoreCanvas has one. So does my solo project.

I set the premise and the voice. Claude drafts the spots in batches, a script casts each one to a synthetic announcer, and the player talks over the last ten seconds of the song the way a real one would.

426

Spots written

3

Stations on air

0

Human announcers

SCORECANVAS

The tool I wanted for eighteen years.

ScoreCanvas is a browser canvas for building adaptive game music: loops, transitions that wait for the bar, covered seams, vertical layers. It is also the course. Students build a score on a node graph, hear every state change as it happens, and turn in the assignment as a URL.

The film was captured from the live product by a script, cut on bar lines and narrated by a synthetic voice. It ends on a full adaptive score built from one melody sung in a car.

10

Lessons on the canvas

3

Companion books

0

Installs required

CATALOG EXPLODER

One finished track in. A whole score out.

Most music is delivered as a stereo file that plays front to back and stops. Games can’t use that. Catalog Exploder takes the stereo file and returns it as a system: stems, a seamless loop and five intensity suites, tempo and key locked to the original, each one checked by ear before it ships.

I built it on a Friday. The domain, the storage, the stem separation and the checkout were live the same day.

This is “Lebec,” a cue from my catalog. The first node is the original. The other four were generated from it. Switch states and the change waits for the next bar line.

THE DAY JOB

A meeting became a manual.

The onboarding for my next game was a 68-minute screen recording of two people explaining their music pipeline. I handed the recording to Claude and asked for a manual I could actually follow.

The same day I had one. Every step is a checkbox. The ones that need a picture get the frame from the video where they happen, and every frame carries a timestamp to jump back to. It is structured like a game, with missions and objectives and progress saved in the browser, because that is how I will use it.

The game is unreleased, so the manual stays private. The method does not.

THE LAB · WEEK 40

Seven engines, one brief.

I keep testing alternatives to the commercial music model my tools depend on. So every engine I could reach got the same brief: five game-score cues, 90 seconds each, 80 BPM, D minor, 4/4. 85 takes went onto a blind ballot.

YuE2 is the interesting one, because it runs on my own machine. It sounds like it belongs in the room. It does not take direction yet. Before it renders, it writes itself a plan, and across ten takes the plan matched the requested tempo once and the key twice.

One useful finding: it sings unless you stop it. An empty lyric sheet produced more than a hundred vocal notes. Section tags with no words produced none on four takes out of five.

SCORECANVAS HOLD · WEEK 40

Hold music that takes requests.

Hold music may be the most-heard and least-loved music there is. So I built a version where the caller is the player. The line opens on an exploration cue at its native tempo. The keypad adds percussion, a melody or a bass pedal, moves to the stealth cue on the bar line, records a lyric, or plays back a tune you whistle.

When the agent picks up, the music hands off to the cue’s own composed ending. A fade is an apology.

The demo page runs the same engine as the phone line. The phone number is next.