Test Driven Songwriting: Joseph Schwartz’s AI Method

A retired corporate technology executive may listen to as many as 100 generated attempts before finding one piece worth keeping. He sets aside 99; one becomes the foundation. The process sounds inefficient until the job becomes clear. Drawing on test-driven programming, Joseph Schwartz does not ask AI to execute a finished creative vision. He starts with a message, tests competing treatments, listens for what works, and allows the strongest result to change his plan.

Generating another track is easy. Hearing why the current one is wrong is harder. Is the cadence fighting the words? Is the vocal too cold? Did the genre flatten the message? Once Schwartz can answer those questions, covers, meta tags, rewritten lyrics, mixing, and mastering become practical tools instead of random controls.

From Hyatt application development to an unplanned music practice

Schwartz came to music through process work. He started as a computer programmer, moved through project management and executive management, and ended his corporate career as VP of application development at Hyatt Hotels. Earlier positions included Andersen Consulting, described as what is now Accenture, along with Verizon and CVS. At Hyatt, he was responsible for customer-facing applications and the customer experience.

The useful carryover was not the code. Schwartz had spent years observing how corporations make decisions, how people complete work, and where the process breaks down. Those concerns eventually entered his lyrics, along with the habits behind them: define the problem, evaluate the tools, run an experiment, inspect the result, and revise the process. After retirement, AI gave him a way to investigate how ordinary tasks might change. Releasing music was not part of the plan.

His first serious music test favored speed over polish. Could he take each day's biggest news story and produce one song a day? The resulting playlist captured a significant period in the news, though Schwartz did not consider it great music or expect people to seek it out. It was a proof of concept. He then began changing lyrics, writing a chorus before building outward, and testing different structures. The medium had changed, but his method had not. He was still experimenting with tools and working out how best to deliver a message.

He was still experimenting with tools and working out how best to deliver a message.

Test-driven programming starts with listening, not notes

Schwartz compares musicians with programmers because both can build upward from the smallest component. A musician may start with notes; a programmer starts with lines of code. Schwartz works from the opposite direction. He listens as an audience member and asks the blunt question that matters when the track plays: "Do I like it?" If the answer is no, he must identify exactly why.

The approach comes from test-driven programming and an agile framework. Rather than spend three years defining every detail before building anything, he makes a piece, examines it, and decides what needs to change. In songwriting, the chorus might need a cleaner cadence, the arrangement less density, or the vocal more anger and less polish. "Make it better" is useless direction. "Shorten this phrase, remove two verses, and make this section more dancey" gives him a specific change to test.

For Schwartz, AI works like an unusually broad group of artists without egos. He can request another interpretation without negotiating schedules or defending the choice. Most results may be disposable. If 99 pieces fail but one contains an interesting vocal treatment, instrument, or structural turn, that is enough. He builds from the useful fragment. The finished track may bear little resemblance to the opening idea because each listening test can reveal a direction he could not see at the start.

How he moves from a message to hundreds of variations

Schwartz now starts with an idea because the music supports the message. He may write a verse and chorus, then use Gemini to explore possible words. The first answer is rarely usable as finished material. He puts down lyrics without laboring over every syllable, establishes enough structure to test the song, and requests around 10 genre or style messages. Those directions may specify certain instruments, a more dance-oriented section, orchestral backing, or virtuoso piano.

Then comes the volume of work that outsiders rarely see. Schwartz may generate outputs hundreds of times before hearing something he wants to develop. Only after choosing the sound does he tighten the writing, replacing a word, repairing a syllable pattern, discarding a phrase, or cutting two verses when the track grows too long. Sound and wording are not separate stages with a tidy handoff. Each requires changes to the other.

He also avoids rebuilding a promising track from scratch. Using covers and meta tags, he can preserve part of an existing result while directing a different treatment. In one example, he kept 40% of the original audio and layered a new style over it. He used that approach on an industrial EDM track with a deliberately cold political tone, placing industrial EDM beneath alt R&B vocals. Neither ingredient alone was in his usual range, but he liked the combination. The feedback loop is precise: test individual parts, retain what works, and replace the rest.

The finished track may bear little resemblance to the opening idea because each listening test can reveal a direction he could not see at the start.

John Henry and Ready for Doom turn the method into an argument

A visit to the Museum of African American Music in Nashville led Schwartz to question a common defensive claim: acknowledging AI's involvement while stressing that a person wrote every lyric. He took the opposite approach and chose a traditional song for an AI-led update. "John Henry" was an obvious candidate because its central conflict already pitted a man against technology. Bruce Springsteen and many others had covered it, generally preserving its traditional character even when they expanded the instrumentation.

Schwartz first retained the lyrics and created a synth-pop version. He liked the concept as commentary, but it did not find a wide audience. So he changed the test. In the rewritten version, John Henry is a musician competing against a tech bro. He plays so hard that he dies with a pick in his hand, then is buried near fiber cables through which listeners hear hybrid music. Schwartz kept the original story's structure and made two versions, one with a male singer and another with a female singer.

"Ready for Doom" applied the same method to current politics. Schwartz connected contemporary political disruptions and talk about the rapture with the Four Horsemen of the Apocalypse. Each verse brought another horseman into a chaotic party before a final dance rant. He used Suno to build from his vocal cadence, though he says he cannot sing, released the track in four genres, and used Audacity to adjust one vocal beyond Suno's original output. The message set the structure. The tools allowed repeated trials until the delivery fit.

The message set the structure.

When 4,000 followers become a creative constraint

Making a track and finding listeners are separate jobs. Schwartz had to learn promotion, advertising, social media, and Spotify analytics. That created a direct conflict. An audience of 4,000 followers might associate him with an '80s rock sound, while his next release could be dance pop with a female vocal. Analytics can show him what the existing audience accepted. They cannot decide whether he wants to make something different.

Because Schwartz does not need the project to make money, he refuses to remain within one genre. The choice makes audience development less efficient, but it preserves his original reason for making music. "Ready for Doom" still triggered Spotify milestones and produced a large increase in listeners. At one point, he recorded 25,000 streams in a day. In a later month, listeners across 80 countries played his songs between two and three times on average. Those numbers moved the project from possible music to viable music. Profitable music remained a separate experiment.

The decline after the peak still bothered him. Numbers have that effect when the highest one becomes the baseline. The useful correction was to consider the full result: hundreds of people in 80 countries had listened, far more reach than Schwartz imagined when he started. His conclusion is more practical than calling the work good or bad. He makes the piece, then looks for its audience. Making money depends partly on finding those people efficiently; resonance does not. Schwartz can keep crossing genres, release music when he chooses, and stay intellectually and creatively engaged without letting Spotify's latest spike dictate his next song.

Making money depends partly on finding those people efficiently; resonance does not.