My guitar teacher introduced me to the BOSS GT-1, a 2016 multi-effects pedal that simulates amps and dozens of effects. It does almost everything. The problem is getting there.
There are hundreds of settings spread across menus, named after gear. I got the logic quickly, because I spend my days designing interfaces. That's exactly why I saw the problem: it was made for people who already think like a sound engineer. If you just want to play a song, you stay on the factory sound.
I didn't want to learn the name of every knob. I wanted to ask for the sound.
Day 1: the machine answers
I installed the driver and the official app on my Mac. Poking around, I found that the app keeps a map of every setting on the pedal in a file. With that map, Claude could read and change any setting over the USB cable.
I asked for a tone. It sent the commands. I played and said: "nothing changed".
It had changed. I checked in a program that shows every message exchanged with the pedal, and the commands were identical to the official app's. The change was just too subtle for someone expecting to hear a difference.
Technically, everything was right. Even so, I was already starting to distrust the tool.
The fix wasn't technical. The first change became impossible to miss, and the tone name started changing on the pedal's display right away. Proof you can see.
People who play don't talk in parameters
None of my requests had a technical name:
"I want the Knee Socks solo." "Needs more drive." "Sounds weird lol." "Turn it down. Neighbors."
Claude knew what a fuzz and a noise gate were. I didn't need to. Its job was to translate what I said into a small adjustment, and tell me what had changed.
When I said "it's dirty", it only touched the noise reduction and left the rest of the tone the way I had approved it. Changing everything would have fixed the noise and ruined the sound I liked.
Day 2: the mistake that became a rule
On the second day, I asked it to save a tone to the pedal's memory. Claude saved over the Iron Man tone I had built the day before.
I had backed up every tone that morning. I got everything back. But the lesson stuck: trusting the AI to be careful isn't enough.
Saving now follows three steps the code won't let you skip: show what's in that slot, wait for a yes, and keep a copy of what was there before.
Volume had the same story. I asked for louder and it came out way too loud. Today no setting goes up more than 15 points at a time, even if someone asks for 40. Ears and neighbors don't have an undo.
Measuring isn't listening
Claude also learned to listen. It records the sound coming out of the pedal over USB and measures volume and noise. In eight seconds, it matched the volume of one tone to another.
I thought it was quiet.
The numbers were right. But a distorted sound doesn't feel as loud as a clean one at the same measurement. Measuring gets you close. The person playing decides.
Twice the problem wasn't even the sound. It was a cable with a bad contact. No AI fixes that, but it can ask about the cable before touching the tone.
From my Mac to other people
By the end of the second day, all of this had become an extension for Claude Desktop. The rules born from mistakes went into the code: volume in small steps, confirmation before deleting anything, backup before saving, undo for everything and error messages that say what to do.
Two weeks later, my teacher tested it with me. We built three tones just by talking: Comfortably Numb, Sunday Bloody Sunday and Rosanna. On the U2 one, his comment was "it's too clean". Claude raised the amp gain and the sound got there. Then it adapted Rosanna for a small 15-watt amp, which is what a lot of people have at home.
The name changed along the way too. The project was called Dial In, an English expression for getting a setting just right. In Portuguese it means nothing and sounds like dialing a phone. It became GT Pilot: you play, it pilots the pedal.
Where it is today
GT Pilot is free and open source. It has a page with 60 songs and the path to each one's tone on the GT-1, from Black Sabbath to Legião Urbana.
Each style also got its own page, telling the same product in the visual language of the people who play it: metal in blackletter, punk in poster collage, reggae in the genre's colors, folk in a calm serif. Nine versions, from rock to folk, each with the tones of its style.
The AI solved the machine side in two days. The rest of the time went into making it work for whoever is on the other side.
2016: the pedal's year. 2 days for the AI to learn to control it. 926 settings per tone. 14 tools in the extension. 1 tone overwritten and recovered. 0 parameters the guitarist needs to know.
For you, designer: the tool only earned trust once the first change became impossible to miss and the tone name showed up on the display. Proof you can see solved what technical precision alone couldn't.
For you, PM: requests came in as "needs more drive" and "turn it down, neighbors", never as parameters. Designing from the way people talk, not from the device manual, is what made the tool work for people who play.
For you, tech: the safety rules (small steps, confirmation, backup, undo) came from real mistakes and went into the code, not the prompt. Trusting the model to be careful isn't enough when a mistake erases someone's work.
