Can You Fix an AI Song Without Regenerating? WAR COMES: Musicfy to EditLab Tracker
Share
When the Song Is Worth Keeping but the Sound Needs Work
We took a real Musicfy development master of WAR COMES into EditLab Tracker. One developer, Claude-assisted editing, four documented passes—and a bigger question than “can AI make another version?”
Sometimes an AI-generated recording has the right idea, voice and attitude but still falls short as a production. The instinct is to regenerate it and hope the next result fixes the mix. Yet another generation can also change the performance you wanted to keep.
That is the problem behind this experiment. I had already been developing WAR COMES through Musicfy as part of the Jack Righteous Creator Lab. I liked the creative direction, but I wanted to know whether targeted production work could improve the recording without sending the entire idea through another round of generation.
I shared the Musicfy development master with Mark Kaminsky, developer of EditLab Tracker. Mark worked on the track with Tracker and Claude, returned three intermediate MP3 passes, a fourth final MP3 and an editable Tracker project, and explained what he did, where the tools helped and where human listening remained essential.
What this is: A documented collaboration on one existing recording—not a blind product comparison, a claim that Musicfy performed the repairs, or a promise that every song will respond the same way.
Hear the actual development chain
Start with the original Musicfy development master, then move through the four EditLab versions in order. Listen at similar playback volume; louder does not automatically mean better. We are retaining every pass because the route to a finished decision matters as much as the ending.
The original development master
The un-repaired Musicfy source sent to Mark. This is the proper “before” reference. The original is already documented in the WAR COMES Creator Lab.
Hear the Musicfy source in the existing test ↗First repair pass
An early exported state. Compare vocal position, low end and overall arrangement with the original before drawing a conclusion.
Open Version 1 audio ↗Second repair pass
A preserved development version. We are not assigning individual fixes to a pass without Mark's exact change log.
Open Version 2 audio ↗Third repair pass
The next version in the chain. Focus on whether each change supports the song's identity rather than simply increasing production complexity.
Open Version 3 audio ↗Mark's final example
The developer's completed demonstration, including vocal, rhythm, bass, harmony and arrangement interventions described below.
Open final Version 4 audio ↗Inspect the working session
Mark also supplied a roughly 470 MB Tracker project so the edits and added notes can be inspected rather than treated as a black box.
Open Tracker project ↗Audio access note: The EditLab version links currently open Mark's original Google Drive files rather than embedded players. The original Musicfy source is available through the linked WAR COMES Creator Lab test. Direct on-page audio playback is planned for a future update.
What the developer actually changed
Mark's account is unusually useful because it names specific musical problems instead of claiming a generic “AI remaster.” His reported interventions span both repair and new production choices.
| Production area | Reported EditLab work |
|---|---|
| Lead vocal | Separated out unwanted vocal fizz, eased a nasal or “honk” resonance around 1.1 kHz, and adjusted vocal level phrase by phrase. |
| Backing vocals | Moved them forward and widened their presentation. |
| Horn / harmony | Reduced a prominent F-sharp overtone against a low D that was creating an unwanted major-third colour in a minor-key passage. |
| Drums | Wrote a new kick, clap and hi-hat pattern around the song's groove. Mark reports that kick placements stayed within three milliseconds of the original beats. |
| Bass | Added a bass-guitar performance following the original deep sub-bass line, with some fills, to help translate on smaller speakers. |
| Arrangement | Opened up the verses, made the hooks hit more fully, added a full stop before the first hook, introduced violin and cello material, used delayed horn stabs, and changed the ending. |
| Levels | Reported a master around the source's loudness without digital clipping. |
These are Mark's descriptions of the completed work, not independently verified stem-by-stem measurements. He describes the new bass, violin and cello recordings as public domain; confirm rights and license documentation for any future commercial release.
How Tracker and Claude were used
Mark's workflow begins with separating the song into editable parts in Tracker. From Tracker's Help menu, he connects Claude using “Use EditLab Tracker with Claude…”. The creator then describes a specific change in natural language while the track remains open for listening and adjustment.
- Keep the generation you believe in. Save an original source and identify what should not change.
- Diagnose a bounded problem. Is the blocker fizz, resonance, bass translation, an awkward arrangement moment or something else?
- Split and work on the relevant part. Use stems, tracks, EQ, levels or editable notes where the material allows it.
- Listen after each meaningful intervention. Measure where measurement helps, but let musical context guide the decision.
- Export and compare the versions. Keep each pass so you can decide whether a repair made the song more convincing—or simply more different.
Claude can help with timing, levels and clipping. It cannot replace the human decision that a particular horn tone feels wrong or that a verse needs less density. The producer still has to direct and judge the work.
Where repair becomes production
Removing fizz and controlling a resonance are targeted repairs. Adding a bass-guitar line, changing drum parts and bringing in strings are broader production decisions. Both can be appropriate, but they should not be described as the same job.
Protect the lyrics, performance identity, mood and sections that already work.
Fix audible problems without discarding the good parts.
Choose new arrangements, instruments and transitions only when they serve the artistic intention.
Mark also identified a real limit: sounds fused together in the original generation may remain difficult to isolate. His example was a horn embedded with a pad—he could reduce an overtone, but not cleanly remove the horn as a separate instrument.
What the four audio files establish—and what they do not
We verified the four provided MP3 exports and their sequence: Versions 1–3 each run about 4:00; Version 4 runs about 4:05. A basic loudness scan showed Version 1 around −15.1 LUFS integrated and Version 4 around −13.8 LUFS. That modest loudness difference is a reason to level-match before judging sound quality—not evidence, on its own, that one version wins.
The files demonstrate that several edit states were preserved and that the developer produced an expanded final arrangement. They do not establish which individual repair first occurred in V1, V2 or V3, how another song would respond, or whether the final should become the official release. Those decisions require further listening, source comparison and creator approval.
The creator's decision: Does the final keep the emotional core of WAR COMES, and is the production improvement worth the added complexity? That is the question we're documenting—not pre-answering.
Try this on your own AI-generated song
Pick one track you already care about. Write down three things you refuse to lose and the single defect you most want fixed. Export the original, work on that defect, then compare a new pass at matched loudness. Only move to rearrangement after the targeted repair is judged. Keep the originals so you can return to the best decision.
If the original problem is generating a song or developing a custom voice, the Musicfy Creator Hub is the right starting point. If the song already exists and needs smaller, controllable edits, explore Mark's earlier vocal-fizz repair demonstration and the EditLab Tracker review.
Production credits and editorial disclosure
Source track and editorial direction: Jack Righteous / Gary Whittaker. Source platform: Musicfy. Developer-led repair and production demonstration: Mark Kaminsky, EditLab. Assisted workflow: EditLab Tracker with Claude.
JackRighteous.com's Musicfy creator journey includes sponsored work. EditLab independently supplied complimentary Pro codes and developer time for testing and education. This article documents one collaborative workflow, including its limitations; the editorial assessment and any eventual release decision remain Jack Righteous's own. There is no guarantee of a similar result in other projects.
