One Sentence at 1 AM
"Make a video."
Our third video. Twenty minutes. Executive audience. A Japanese precision machinery maker's mid-term plan. Eight visual styles to choose from—hand-drawn watercolor. Forty-eight film crew roles, twenty-four think-tank analysts on standby.
Eight hours later, a 15-minute-50-second video is live on YouTube.
But this is not a "we made a video" post. It's a "the real output of those eight hours was six reusable tools" post. The video is a one-time deliverable. The skills are permanent.
1. Deliberate Absence: Learning Not to Fill
Six boxes on a strategy diagram. No arrows between them.
Not an oversight. Deliberate. Slide 26 has a 3-second pause for the same reason. The blank space is the content—it says "we're not connecting these dots for you, you decide."
But automation pipelines don't think that way. Phase gates check that every node has a connection. Prettier removes extra line breaks. AI autocomplete sees a gap and wants to fill it.
So we built deliberate-absence: max 3 seconds, explicitly tagged, pipelines configured to route around it.
What we learned: Tools instinctively erase empty space. If you need it, give it a name so the tools recognize it.
2. Batch Image Gen: Stop Babysitting 48 Images
GPT Image-2 takes 60-120 seconds per image. Forty-eight images in serial: 1.5 hours.
A for-loop got killed by shell timeout three times. Each restart meant starting from scratch—because we couldn't tell which ones succeeded.
We wrote batch-gen.js: 3 parallel workers, 180-second per-image timeout, auto-retry on failure, progress written to a manifest. Next time we need 48 slides: one command, go get coffee, done.
What we learned: A for-loop is not a batch job. Real batching needs retries, parallelism, and independent failure domains.
3. ASR Subtitle Alignment: When Math Lies
The video was sped up 1.25x. We multiplied all subtitle timestamps by SPEED=0.8. Wrong. 1.25x speedup is not linear across 347 subtitle segments. Each one drifted a little. By the end, everything was off by over ten seconds.
We tried multiple multipliers. None worked.
Finally we downloaded the YouTube audio track and ran Whisper medium. Three minutes of transcription. 347 segments, perfectly aligned—because they came from the actual speech waveform, not a math formula.
"Use ASR, not multiplication."
What we learned: When you have the raw signal (the audio), don't approximate it with a derived formula (speed multiplier). Go back to the source.
4. TTS Shootout: Do This at Phase 0
Six TTS engines: Kore, VoAI, Edge-TTS HsiaoChen, HsiaoYu, and two others. Six rounds of testing to discover:
- VoAI butchers Japanese loanwords (like the company name)
- Kore has a 100 calls/day quota—not enough for long videos
- HsiaoChen is free, unlimited, and passed on the first try
But we ran these tests at Phase 3. If we'd done this at Phase 0—when choosing the voice actor—we wouldn't have hit a wall mid-production.
What we learned: Picking a voice isn't "choose one that sounds nice." It's "run a shootout, eliminate anyone who mispronounces domain terms, hits quota limits, or has the wrong prosody pattern." Do it before you start.
5. Concat Audio Check: The Invisible 44100Hz
Forty-eight audio segments. ffmpeg concat with -c copy. Segment S26 was a silent track at 44100Hz. The other 47 were 24000Hz. -c copy doesn't convert—from the first mismatch onward, everything went silent.
This is a bug you only catch by listening to the entire rendered video. The timestamp was ticking. No sound.
Now we have a preflight check script: scan all audio segments for sample rate before concat. Mismatch → alert.
What we learned: -c copy is fast, but it's a blind concat. Production pipelines need a preflight check.
6. Hardsub Burn: The Day Homebrew Betrayed Me
Homebrew's ffmpeg doesn't include freetype.
No freetype = no drawtext = no subtitles filter = no ASS filter.
Three hundred forty-seven subtitle segments waiting to be burned into the video.
Downloaded a standalone ffmpeg binary from evermeet. One command: speed change, watermark, 347 subtitle burns. That binary now lives in our tools directory. Homebrew's version can keep being the backup. But when it's time to burn subtitles, it doesn't exist.
What we learned: Package managers install general-purpose builds. Specialized workflows need specialized binaries. Put them in a fixed path, give them a name, document it.
The Real Output
"Make a video."
A 15-minute-50-second video went live on YouTube. That was the deliverable.
But what those eight hours really produced were six skills, six scripts, six lessons. Next time someone says "make a video," we don't start from zero. batch-gen.js generates the slides. asr-align locks the subtitles. concat-check scans the sample rates. ffmpeg-full burns the hardsubs in the background.
The real output of a video is not the video. It's every pitfall you'll never fall into again.
Compounding isn't getting it right once. It's turning every mistake into the starting line for the next round.
Top comments (0)