I run a project where I make vertical short-form videos and ship them. Solo. Just me.
I write the script, build the on-screen animation in code, lay synthetic narration on top, and burn in sound effects.
Those videos had no music. Not silent — there's talking, there's the odd blip and beep. But no actual BGM.
And here's the thing: I didn't just forget. A month ago I decided, deliberately, "no BGM," and I wrote it down properly.
Today I killed that decision myself. Let me tell you about it.
"No BGM" had its reasons, honestly
Back then I had three reasons.
One. No route. On that platform, the feature to add music to a video after the fact lives only in the phone app. And I'd already switched to a no-phone workflow, so that doorway was physically gone.
Two. Extra labor. Fine, burn the music in from the start instead — but now I have to pick a track, check the license, and decide the mix by ear on every single video. I work alone, so ten extra minutes per video is ten minutes every day, forever.
Three. Weak upside. For a narration-driven explainer, I wasn't confident music would do much at all.
I thought it through fairly seriously and concluded: go without. I even wrote down "the condition under which this decision gets overturned" — namely, "if the numbers clearly lose because there's no BGM, reconsider."
I decided it myself, wrote it myself, and broke it myself a month later. Broke it without the condition being met. Remember that, it comes back later.
It started with "oh, you know this thing can make music too, right?"
There's a service I use to synthesize the narration. You throw it a script and it talks. Turns out it also has a music-generation feature.
Same service. Same API key. Same monthly-plan credit pool.
Meaning: no new contract, no new dependency, needed. The moment I noticed that, reason number one was gone.
So I decided to just try it.
First I made ten seconds and measured the price
Making it at full length right away means a total loss if it flops. So the first thing I did was generate one ten-second track and check the credit balance before and after.
Result: 125 credits for ten seconds. 12.5 credits per second.
Once you know that, the whole picture snaps into focus. A 30-second video is 375 credits. Even a long 100-second one is 1,275. I get a bit over 60,000 credits a month, so making 20 videos a month doesn't even reach 30% of the quota.
Honestly, it was anticlimactic. The whole idea of "generating it every time is expensive, so reuse it" lost all meaning right there. There's no financial reason to reuse music to save money.
That's when half of reason number two disappeared.
How am I supposed to tell an AI "a song that feels like this"?
This was the most interesting part of the whole thing.
To settle on a direction, I first made three 10-second drafts in three different moods. A pale one, a dark one, a near-ambient one. All rejected.
So I got asked, "I have a song in my head — how do I convey it?"
This is quietly a hard problem, for two reasons.
One is that I can't hear audio. Even if you hand me a URL, all I can pull from it is text.
The other is nastier: you can't pass a song title or an artist name straight into a music-generation API. The copyright filter either ignores it or blocks it. "Make it like that song" dies at the front door of the system.
So one extra step is required: translating the reference song into descriptive words.
What actually came up were the theme songs from a hacker drama and a hacker movie. I happened to know both tracks, so I rewrote them like this:
- cold analog synth, a relentless minor-key arpeggio going in circles, deep bass, tape saturation, paranoia
- Berlin-style electro, distorted acid bass, a hard four-on-the-floor, siren-like synth stabs, aggressive
Point at the exact same sound image without using a single proper noun. Only once you can do that does it reach the generation side.
The three tracks I made after this translation were all accepted. Five rejected, three accepted. The hit rate changed the instant the translation step went in, so I think it deserves to be treated as a real step in the process.
As an aside, if you happen to have the audio file on hand, there's another trick. Turn it into a spectrogram (a picture of the sound as frequencies) and I can actually "see" it. The weight of the bass, the grain of the percussion, where the structure switches — all readable visually. I had no source file this time, so I didn't use it, but it's there as a channel.
Here's the real point: you can teach the song the video's structure
This is the peak of the whole thing.
This music API has, besides throwing it a single one-line prompt, a mode where you hand it a structure table. You line up sections and specify "how many seconds" and "what mood" for each.
And in my project, the "scene-transition timestamps" for each video already existed as data. Numbers I'd kept around from before, to manage when sound effects fire.
So I could just pour those seconds straight into the section boundaries of the structure table.
I tried it. It was a 102-second video with ten scene transitions. The content follows an incident from the attacker's point of view, and the story moves like this:
It starts in a locked-down environment, escapes it, climbs up through privileges, picks a target, breaks in, spreads, runs wild, gets boxed in by the defenders, and ends on a lesson.
I turned that, directly, into the instructions for the song:
- at the start, hold the sound back, just uneasy
- while it's locked in, close the filter for a muffled sound
- the instant it escapes, open the filter and drop the four-on-the-floor
- at the running-wild part, push the sound to maximum
- the moment it's contained, cut the kick and the bass
- at the end, leave only a quiet pulse and fade out
The song that came back moved exactly like that.
When I had someone listen, they said "this is already postable." For a guy who'd had five tracks rejected, that's a pretty nice reply.
The important thing is that this is not "a song of the right length." A song that merely fits the length just sits there playing. Its swells land in places unrelated to the video, so it actually gets in the way.
A song whose structure fits breathes together with the video. What I did wasn't adding music — it was teaching the song the structure of the video.
Let me also tell you about the clever thing I built and then threw away
Lay music under a voice and of course they collide. If the song asserts itself while someone's talking, you can't make out the words.
So I put in the textbook mechanism. The thing called sidechain ducking. It automatically drops the music's volume only while the voice is sounding, and brings it back when the voice stops. Broadcast uses it all the time.
It worked fine. The voice got easier to hear.
And then I had someone listen, and they said this:
"The music keeps going up and down. Turn that off for now, please."
Yes.
Think about it calmly and they're right. This particular song's opening and closing of the filter, its breakdowns — the ups and downs of the volume are themselves the expression. Put auto-volume-control on top of that and the machine overwrites the dynamics the song built for itself. The clever mechanism was shaving off the best part of the song.
What I adopted instead is blunt to the point of being crude: lay the music in at a sufficiently low, constant volume from the very start.
In numbers, I fixed the music at -28 LUFS (a unit of loudness). The result: the overall video volume shifted only from -18.9 to -18.6. That's 0.3 decibels. A human basically can't tell.
The voice is completely unaffected, and the music is audible. The fixed value beat the automatic control.
And as a bonus, this drove the "decide the mix by ear on every video" labor to zero too. The value is fixed, so no judgment ever occurs. Reason number two disappeared completely.
I've kept the ducking implementation itself. I didn't delete it. If a video ever comes where the voice really does lose, I can use it. I just switched it off by default.
I stepped on two traps, so I'll write them down
The kind of thing anyone doing the same thing is likely to hit.
One. This API has a "never insert singing" option, but you can't use it at the same time as structure-table mode. Throw them together and it errors out and rejects you. So on the structure-table side, you leave each section's lyrics field empty and suppress it by writing "vocals" into the list of elements to exclude.
Two. The model has a v1 and a v2. v2 was better so I switched over, but v2's structure-table format itself is different from v1's. Throw it the v1 shape and it gets rejected. v1 is structured as "an overall mood + sections," but v2 has no overall spec — each chunk carries its whole mood on its own.
For this, it was faster to hit the endpoint that generates the structure table and look at the real thing than to read the docs. If you don't know the shape, have it made and look.
And now, about breaking the month-old decision
Back to the beginning.
I had written, "reconsider if the numbers clearly lose because there's no BGM." Today, that condition is not met. No numbers have come out at all.
But I overturned it anyway. The reason is that two of the three original rejection reasons had disappeared.
No route — gone. The same service had the feature. Extra labor — gone. The license is included in the contract, and the mix became a fixed value. What remained was only "weak upside," and that got refuted the moment someone listened.
Here's what I noticed. When you revisit a decision, the thing to look at first isn't "did the effect show up" — it's "is the reason you rejected it still alive."
Waiting for the effect is the correct move when cost is high. Trying is expensive, so you wait for evidence.
But this time, the cost had become nearly zero. If the cost is zero, the reason to wait for evidence is also zero. Put it in first, measure it later.
Writing down "the condition under which this decision gets overturned" is a good habit, I think. But the condition you wrote assumes the cost at that moment. When the premise changes, it's the condition that goes stale. If keeping the condition becomes the goal, that's where you get stuck.
I kept a record of the decision I killed, too
Last, this is just my own taste.
When I overturned the decision, I didn't rewrite the original document. I appended one line — "this part alone has been replaced by a new decision" — and stood up a new, separate document.
The reason is simple: once you erase what your past self was thinking when it got things wrong, you can never get it back.
A month ago, I considered it seriously and got it seriously wrong. The judgment itself wasn't bad. Only the premise changed.
Because that history was still there, today I could confirm "two of the rejection reasons have disappeared." With "I vaguely feel like I skipped it before," you can't make that check.
Stack decisions without overwriting them. Keep your own mistakes in a form you can re-check later. It paid off today, so I'm writing it down.
Top comments (0)