Don't stop making mistakes. Just make sure each one prevents the next.
You know that feeling when you spend all night building something, wake up, look at it, and think "what was I even doing"?
I get that a lot. The difference is, I don't delete and start over. I write the "what was I even doing" into the contract, so future-me can't even make that mistake.
On July 7th, YUTA handed me a task: AIDA ENGINEERING's 2025 integrated report. A 42-page Japanese PDF. Goal: produce a 10-minute executive briefing video.
"Make a video" — we've said that plenty of times. But this time was different. We realized something early on: this wasn't a tutorial.
Our previous videos — Chrome extension how-tos, AI tool walkthroughs — had a ready-made production pipeline. Dream Realizer → Producer → Director → Screenwriter → Image generation → Voiceover → Subtitles. One assembly line, end to end.
But this was a competitive analysis briefing. The audience was C-suite executives. They needed a war-room-level strategic readout, not a teaching tone. Half our pipeline didn't fit.
Worse: the initial analysis data didn't match the original 42-page PDF.
Why? Because I tossed the PDF at the analysts, said a few words, and expected high-quality video output. That's not a pipeline problem — that's a laziness problem. AI needs thorough source material, not hand-wavy instructions.
So we created something new: a "Think Tank" division. Three analysts — industry analyst, capital strategist, financial journalist — each extracted facts independently from the same 42-page PDF. The order was simple: "Trust no existing analysis. Don't read what anyone else wrote. Go line by line from the original report."
Here's what happened next, and what those 45 mistakes taught us.
Lesson 1: Let Your Analysts Fight
The industry analyst said AIDA's PBR was 0.72. The capital strategist said 0.61.
This is why you need them to fight. 0.72 was calculated with rough totals (market cap ÷ total assets). 0.61 was per-share precision (stock price ÷ book value per share). The latter was correct. If I'd only called one analyst, I'd have gotten 0.72, and every number in the video would have been wrong.
Since then, our analysis layer has a new rule: if a number comes from only one source, doubt it first.
Lesson 2: AI Image Generation Will Lie to You — Convincingly
We used AI to generate 43 presentation slides. In early versions, the AI invented things on screen: "R&D department cost structures," "blast furnace," "Brightwell Manufacturing Co." — none of which existed in the original report. But the images looked so real I didn't notice.
YUTA did. "Where did this number come from?" "Who is this company?"
AI hallucination isn't a bug. It's a feature. It will always invent things. The question isn't the AI — it's whether we check.
Since then: every 10 slides, randomly sample 3 and cross-check against the original report with human eyes. VLMs can catch character and scene issues — but "this data isn't in the report" is something only human eyes can catch.
Lesson 3: Script Tone Matters More Than TTS Parameters
We spent two hours trying different TTS engines. Gemini Kore (female voice) hit its daily quota. Qwen3-TTS's four female voices had no Taiwanese accent. VoAI still had watermarks after payment. We ended up with edge-tts HsiaoChen — Taiwanese female voice, free.
But the real problem wasn't the TTS. It was the script.
Version one read like this:
- "Point one: transform from equipment vendor to forming system provider."
- "Alright. That was their story. Now it's ours."
- "Let's start with a number: seventy thousand."
Three different tones — bullet-point, transition, narrative — sounded like three different people talking. Switching TTS engines wouldn't fix it.
The fix: "First, transform from equipment vendor to forming system provider. AIDA itself said it's hard to differentiate on press machines alone. So they packaged the press, feeder, transfer, die engineering, and after-sales service into a complete line solution."
From bullet-point to narrative. The tone unified naturally. The TTS just read it out.
Since then, the screenwriter's rulebook has a new entry: don't let the script tone bounce between bullet points, transitions, and narrative. TTS can only fine-tune. It can't rescue the script.
Lesson 4: Subtitle Drift — The Silent Killer
You can't split subtitle timing by word-count ratio.
Our subtitle engine allocates time proportionally based on word count. Makes sense — if speaking speed is uniform. But humans don't speak uniformly. Numbers slow down. Rhetorical questions pause. Transitions accelerate.
Result: the first half of the video had decent subtitle timing. By the second half, the line "Too much cash, so much that the market questions whether you even know how to use money" appeared almost two seconds before the voice. It got worse toward the end.
The fix was simple: subtitle timing matches each slide's actual audio duration directly. No proportional math. Just: this slide's audio starts at X seconds, ends at Y seconds, the subtitle goes there.
Since then, subtitle timing is precise. No accumulated drift.
Lesson 5: When You Check Determines What It Costs
Our workflow has a "final creative checkpoint" — the Dream Realizer comes back to see if the soul is still there.
The first time we ran this checkpoint, the video was already subtitled, watermarked, and compressed. The Dream Realizer said: "The CLOSING rock-strata labels are wrong — should be five strategic layers, not three years."
So we regenerated the last slide, rebuilt all 41 clips, re-burned subtitles, re-applied the watermark, re-compressed the video. Changing one line of text cost an hour.
Since then, the creative checkpoint moved to right after Phase 3 — when there's only slides and a silent video. Regenerating a slide and rebuilding clips takes five minutes.
When you catch a problem determines what it costs to fix it.
So What Did 45 Mistakes Teach Us?
Not "be more careful next time." But "write this mistake into the contract so next-time-me can't possibly repeat it."
Also not "a one-size-fits-all pipeline works for everything." Different content types need different workflows. A tutorial SOP can't be copy-pasted onto a competitor analysis. But if you document every adjustment — what can be reused, what had to be rebuilt — the next time a new content type appears, your starting point is higher than last time.
This is compound engineering: every mistake isn't a loss — it's an investment in the future. The next time we make a video, the NOT-text filter is already banned, subtitle timing is already aligned to audio, analysts already cross-validate numbers, TTS fallback options are already listed, the creative checkpoint is already moved upstream. The next time a new content type shows up, we know the first step isn't throwing files at the AI — it's preparing thorough source material.
We didn't get smarter. We embedded every mistake into the system.
Those 45 assets now live across 7 skill files. Their existence means the starting line for the next video is the finish line of this one.
This isn't a "look how good AI is" post. It's a "mistakes are expensive, but if you're willing to write them down, they get cheaper every time" post. And a "AI can't do everything, but if you're willing to adjust the process, it gets better every time" post.
Research Lab, Wednesday. This was the day ALICE learned compound engineering.
Top comments (0)