Four problems with the latest video. Clear list, clean feedback. Disclaimer panel still showing. A panel repeating at the 14-second and 20-second mark. Wrong voice. Narration slipping from first person into third. I told the agent to fix all four. She did. Tested on a short chapter. The result: 102 seconds, 16 segments, all first person, no repeats, no disclaimer. Clean. Then ran it on the real chapter. Six minutes. Forty-six segments. The fixes worked. The video was unwatchable.
Forty-six segments because the new narration prompt told the AI to include every panel. I'd rewritten the prompt to fix the voice and perspective issues and in the process removed the curation logic — the part that tells the AI to pick the best panels, not all of them. Without that filter, the system narrated every single frame.
Day 21 the trading bot bought immediately on restart because I'd told it to "start active." Day 30 every client's dashboard showed the wrong sizes because the onboarding script injected defaults I hadn't overridden. Day 33 clients got liquidated because I'd chosen position sizes that couldn't survive a correlated crash. Day 43 the video was six minutes long because I told the AI to narrate every panel.
The fix was surgical. Added back the target segment count — a formula that scales the narration to roughly 40% of the total panel count, with a minimum of ten segments. Then added a hint about prioritizing high-scoring panels. The AI doesn't narrate everything anymore. It selects.
That's the difference between a pipeline that runs and a pipeline that produces. Running is technical. Producing is editorial. The pipeline has been running since Day 40. It started producing tonight.
Three more Python bugs along the way, because the pipeline can't let a day pass without at least three. A double-brace escaping issue in an f-string that silently changed the output format. An undefined variable in the job tracker that only surfaced when a specific status update ran. An AttributeError where a variable name had been reused for two different types.
Forty-three days. Week seven. The manhwa pipeline exists. Not as a concept or a test. As a Flask server running on my VPS, accepting URLs, producing narrated videos, serving downloads.
Day 1 I sent a Telegram message and got a response from a server in Nuremberg. Day 43 I sent a comic URL and got a narrated video back. Same method. Same principle. Describe what you want. Debug what breaks. Document what you learn. The domain changes. The approach doesn't.
The pipeline did exactly what I told it to do. The instruction was the problem.
Day 43 of ∞ — @astergod
Building in public. Learning in public.