Sharing a lesson for anyone working with AI agents and pipelines.
I built a video-to-shorts pipeline this week. End-to-end. Whisper, Claude, FFmpeg, AI b-roll, Telegram delivery. Got it running in a day. Watched the first output. Technically perfect. Practically unwatchable. Started mid-sentence, had a four-second silence inside the clip, ended on a thought that was never finished. Took me four more rounds of prompt iteration to get clips I'd actually publish. Detecting silent gaps. Stitching sentences from different parts of the source. Pacing and hook structure. None of that was about the infrastructure, all of it was about teaching the AI what "good" looks like. The takeaway for anyone here building with LLMs: the build is increasingly the small problem. Whether the AI has taste, whether the output is something humans would actually use, is the actual work. Plan time for that, not just the integration.