# TAKO and AI agent production conversation These excerpts show the production exchanges that materially changed the 127.77-second, 23-shot main film. The wording has been lightly edited for readability. ## 1. Use moving vertical footage, not a single extracted frame ### TAKO > Are you going to take just one frame from the video I gave you and turn that into a video? ### AI agent The image in the Excel storyboard is only a representative frame for composition review. The final work uses the outer-fence and entrance-fence MOV files as continuous live footage, preserving camera movement, vegetation, wind, birds, and other ambience. The vertical footage is not stretched. It remains centered, with enlarged blurred copies filling the sides of the 16:9 frame. ## 2. Match TAKO’s size numerically, not by intuition ### TAKO > Keep TAKO’s size consistent throughout. He is about 160 cm tall. The fence is 140 cm, and the tape above it is at 2 m. You know that, right? He is much too large here. ### AI agent TAKO at about 160 cm, the inner fence at 140 cm, and the gold tape at about 200 cm became a shared measuring system. TAKO was placed on the same ground plane and at the same depth as the comparison object. Against a 140 cm fence, he should appear only about 14% taller. Because image generation made the scale drift again, TAKO’s position and pixel height were locked with local compositing after the real background was fixed. ## 3. Remove abstract arrows and follow the real fences ### TAKO > The orange line does not match the fence, so it is hard to understand. The arrowhead looks strange. The arrow for the outer fence is unclear too. ### AI agent Generic arrows were removed. The inner fence was traced in orange and the outer fence in red, directly along the real posts and tape. Perspective followed the actual ground and fence tops instead of a convenient graphic direction. Exact text, color lines, and measurements were fixed by deterministic local editing rather than generation. ## 4. A usable clip can still be omitted if it weakens this story ### TAKO > I added `シカ再び.mp4`. > > For the final gag, TAKO looks at the outer fence and says, “It’s like Wall Maria.” Then a colossal character walks toward us from far away. After reviewing the sequence, TAKO decided: > We do not need the returning-deer scene in this one. We can use it in another episode. The colossal figure should approach with a deep “DOKO… DOKO…” sound, right? ### AI agent `シカ再び.mp4` was retained but omitted from this final film. The ending was narrowed to a short causal chain: TAKO compares the fence to Wall Maria, notices a deep footfall, and an original colossal figure takes two steps from the same distant background before an immediate cut to black. ## 5. Confirm the interpretation before editing ### TAKO > Hey, do not start making changes on your own. TAKO also needs to remain in the shot with the colossal figure, or the cut will feel wrong. The orange line also slopes the wrong way in the distance. ### AI agent Work stopped while the interpretation was clarified. The next shot would keep the same TAKO in the same framing, position, and scale. The inner-fence line would follow the actual perspective. All requested changes would be confirmed together before implementation resumed. ## 6. Choose the creator-like voice, then fit the visuals to it ### TAKO > Could we create the voice from my own voice data? After listening to test versions: > The expressive one sounds good. Also, isn’t this part much too fast? Later: > I want TAKO to keep using this voice in future episodes. Can we establish a repeatable method? ### AI agent Chatterbox TTS 0.1.7 used a short, roughly 20-second private voice reference for zero-shot speech synthesis. Restrained, natural, and expressive variants were compared, and the creator selected the expressive profile. All 20 lines were then produced without forcing them into three-second shots. Visual duration was extended to at least `voice duration + 0.55 seconds`, and every line remained at 1.0× speed. When five lines needed pronunciation corrections, the approved voice for the other fifteen lines was preserved. Display text and spoken input were separated, and only the affected lines were regenerated. The original voice reference stayed in the private local environment and was not copied into the deliverables. ## What this conversation taught us Natural live-action integration depends on more than visual style. The source must be inspected over time, while aspect ratio, ground plane, depth, real-world measurements, and continuity are fixed one by one. Natural speech should be generated first and given enough screen time. More material is not always better: protecting the central idea sometimes means saving a good clip for another film.