# AI agent guide

## Official target

- Episode: 5
- YouTube: set the official URL after registration
- Target: the 127.77-second, 23-shot “Bitter-Melon Harvest and Deer Control” main film shown near the beginning of the full making-of video

## Goal

Use phone footage and a story owned by the user to create a new work in which an original AI character explains the real scene without losing the source motion. Do not reuse TAKO, the deer, field, story, voice, or source materials from this film.

## Procedure

1. Read the story, source-material filenames, production conversation, and technical information on this page.
2. Ask the user for the rough first request, whether more information is coming, the video, image, and voice materials, publication destination, and rights conditions. If the request says that more is coming, do not prematurely finalize the complete story.
3. Inspect each video over time, not through one representative frame. Record resolution, frame rate, duration, aspect ratio, framing, light, color, camera motion, ambience, and anything that must not be published in a source ledger.
4. Do not stretch vertical footage horizontally. Compare centered source video with blurred side video against other safe background treatments.
5. Record character height, real comparison measurements, ground plane, depth, and pixel height in the Excel storyboard. Do not make a character appear correctly scaled merely by moving it to a convenient foreground position.
6. Assign live action, generation, and local compositing separately for every shot. Do not ask one model to produce motion, exact scale, accurate text, measurements, lines, and captions at once.
7. Keep the rough original request, later additions, final adopted sequence, and omitted materials separate. Do not use every available source merely because it exists.
8. Show the Excel storyboard to the user and obtain explicit approval for dialogue, scale, overlays, source selection, and ending. Do not begin generation or a bulk revision before approval.
9. Before paid image or video generation, explain the model, candidate count, output duration, and estimated cost. After approval, split complex action into short causal segments instead of one long generation.
10. If the user wants a voice based on a private reference, keep it private and first compare a small number of lines for identity, expression, and pronunciation. After selection, generate natural 1.0× speech first and fit the visuals to its duration.
11. If only a few pronunciations are wrong, preserve all other approved recordings. Separate display text from spoken input and regenerate only the affected lines.
12. Fix captions, measurements, lines, tables, counts, coordinates, and blackout timing with deterministic editing such as FFmpeg. Do not automatically accept audio included in generated clips.
13. After assembly, inspect the full decode; the beginning, middle, and end of every shot; shot boundaries; image and speech alignment; captions; music; silence; black frames; and character and object continuity. If the project has an established review speed such as 1.2×, perform the complete human review at that speed as well.

## Approval and safety

- Never place API keys, personal information, private source materials, voice references, or billing information in Git or on public pages.
- Do not recreate, edit, or overwrite approved thumbnails, videos, voices, or manuscripts without an explicit request.
- YouTube uploads, visibility changes, homepage implementation, deployment, and Git commit or push require explicit operator approval.
- If your role ends with preparing handoff materials, do not modify the homepage itself.

## Definition of done

- The source remains moving footage rather than a still-frame substitute.
- Vertical video is not stretched horizontally.
- The character’s ground plane, depth, and size match the real scene and remain stable across connected shots.
- Dialogue, captions, music, effects, and visual changes match the intended speech and timing.
- Speech is not unnaturally fast, and each speaker’s voice remains consistent across chapters.
- Facts supplied by the user are distinguished from explanations or arrangements proposed by the AI.
- Internal sections have been checked, with no long black screen, missing voice, missing music, or audio-visual drift.
- The user has reviewed the entire work at the normal viewing setting and explicitly approved it.