Summary
Keith AI documents a full attempt to replace Descript — a $600/year video editing subscription — by building a custom AI-powered editor using Kimi K3. The project targets eight specific editing capabilities: camera movement, word-level emphasis, sound design, script generation with angle options, video repurposing, filler word removal, and dynamic layout selection. The build is connected to a 500-sound MCP library, and the editing rules are encoded as a reusable “taste file” that persists across sessions.
The video is honest about where things broke. First drafts produced flat animations, mistimed sounds, and what the creator calls “AI slop” when given vague feedback like “make it more engaging.” The fix required highly specific direction — animate one word at a time, use different animations for different concepts, trigger sounds only on specific visual events. Keith documents the iterative revision process (4–20 rounds per clip) and shows how corrections are saved so the editor learns without starting from scratch.
The final section runs the same footage and editing rules through ChatGPT, Claude, and Kimi to test the creator’s hypothesis: that the taste file matters more than which model you use. Results, costs, and time-to-completion are compared. The complete workflow and taste files are offered to viewers for reuse. This is a practical guide for creators exploring AI-assisted video production pipelines, with a real cost-versus-quality analysis grounding the experiment.
📺 Source: Keith AI · Published September 14, 2026
🏷️ Format: Hands On Build
