video-use turns raw footage into finished cuts

This is a smart application of the coding-agent pattern: keep the model on text, timestamps, and edit decisions, then let deterministic tools handle render work.

• –The text-first pipeline is the real idea here: transcript packing, timestamp alignment, and selective visual checks keep the LLM out of frame soup.
• –ffmpeg is the right execution layer because it makes cuts, fades, subtitles, and renders reproducible rather than “AI-style” approximate.
• –The self-eval loop is important; video tools fail loudly on audio pops, subtitle drift, and awkward jump cuts, so automated verification matters.
• –This is best viewed as an AI assistant for rough and intermediate cuts, not a replacement for a full non-linear editor when finishing work gets complex.
• –The local/open-source angle matters: creators and small teams can use agentic editing without handing raw footage to a cloud SaaS.