Overview
A lot of my coding has been related to my music project (I make music as 'MYRY', designed my website, visuals including artwork and videos, manage my own ads and analytics, and built my own data aggregation platform for this).
I became frustrated by the fragmentation of the software available to analyse and caption audio, create audio-reactive visuals or any form of engaging visual content, and separately having to record and edit video, so I built my own.
It started with a basic video editing tool that analyses the audio, transcribes into captions, and separately analyses the stems to create visuals that respond to each element separately.
Through endless iterations with multiple models and rejected imperfect renders, it evolved into a complex software able to edit several GBs of professionally shot footage either manually or entirely through AI, whilst I retain 100% control and oversight of the final output, as the project is created in my native application.
Process
1. The first real design problem
In October the app had plenty of power and a layout to match. A dense left rail with ten or so entries, panels that duplicated each other, a timeline below the fold, and a welcome card that ate the screen. It was an engineer's tool.
I ran seven documented review passes on 12 October, treating the app like a stranger would. The biggest structural change came out of them: song-first became video-first. The editor used to assume a music track defined time. A real project clock replaced that, so a song became optional. That single change unlocked the rest: source bins, snapping, dissolves, independent audio clips, and a timeline that behaved the way editors expect.
2. Making it feel easy without making it weak
Two references pulled in opposite directions. CapCut is easy: one preview in the centre, tools on the left, settings on the right, timeline across the bottom. Premiere is powerful: everything is there if you know where to look.
The answer was a layered design. On 14 October the layout moved to the CapCut shape. Then in late October the Video Studio redesign cut the left rail to six tabs (Media, Audio, Text, Effects, Transitions, Color), moved everything else into a More menu, and gave the timeline a compact icon toolbar. The Inspector kept the pro controls: transform, blend modes, chroma key, masks, colour curves, speed. They simply stay folded until needed.
The rule I kept was that nothing disappears. The redesign has a written map of where every old control went. The original Classic layout stays untouched, one click away, for anyone who prefers it. Both interfaces are views of the same project engine, so switching never costs anything.
Graphics also became real clips on the timeline, with trim, split, hide and delete like any other clip. Audio-reactive visuals and a "music to code" editor gave the app a creative side beyond cutting.
3. The turn toward AI, with a constraint
In November the question changed from "how do I edit?" to "can the app make a first cut for me?" The temptation with AI is to deliver a finished video. I made the opposite rule: AI output must arrive as an ordinary ARV project. Every clip, trim, grade and caption stays editable, and an AI cut is only a proposal, approved as one undoable step.
That became a small production team, each role written down. A producer checks the footage and rights. A music supervisor measures the actual beat. A director writes a treatment. An editor cuts to the music and avoids repeating shots. A colourist keeps skin tone and grade steady. An independent reviewer judges continuity and arc on the real frames. A style study of reference music videos fed the pacing, and trained audio models (stem separation, beat and key detection) made the music analysis more reliable.
4. What the work taught me
The clearest lesson came from a real artist's music video. The first cuts landed on a regular beat grid and the artist felt they missed the bass. Measuring the music showed the bass actually hits off the downbeat. Cutting on the measured hits fixed it. The model's score had looked fine; the artist's ear was right. After that, the rule became: a successful export is not the same as an artist saying yes. Failed reviews are kept, and the team has to open the delivered project, make a change in a copy, undo, save and reopen it before calling it done.
The video went through about 26 numbered versions. Each one is a project the artist can open and keep editing.
Impact
Principles I'd carry forward
- Easy first, powerful on demand. Show six things; keep two hundred one click away.
- Never trade away control for convenience. AI proposes, the human decides, and everything stays editable.
- Local and private by default. The user's footage never needs to leave their machine.
- The person has the final say. Measure the music, then check it against the artist's ear.
- Be honest about what is proven. Reviews so far are internal, and several screens were tested with synthetic clips.