# AI agent guide

## Official target

- Episode: 6
- YouTube: `https://youtu.be/HRmKSoSn3WE`
- Target: the 26-second “Mount Fuji and Its Reflection” opening example and the production activity that created and reviewed more than 50 Aha and spot-the-difference trials in half a day

## Goal

Use subject matter that the user has the right to use to create short gradual-change videos. The AI agent proposes candidates, develops images, renders videos, and performs technical checks. A person returns OK or NG decisions with reasons and makes the final creative judgment.

Do not reuse the original Mount Fuji, cat, dog, landmark, image, video, answer, character, or voice assets.

## Procedure

1. Read the story, input-material filenames, production conversation, and technical information on this page.
2. Ask the user for the desired subject, intended change, elements that must remain fixed, answer description, publication destination, and rights conditions. If the user says that more instructions are coming or asks you to wait for a completion signal, do not begin research, production, or generation until that signal arrives.
3. If the request contains multiple subjects, propose only a small first set in text. Check that neither the subject nor the change mechanism duplicates earlier accepted work.
4. The standard workflow obtains human approval of an Excel storyboard before video production. Only when many short trials with one shared specification are more appropriate, explain the no-storyboard special rule, its scope, and its quality gates, then obtain explicit approval.
5. Before any paid image generation or other paid operation, explain the process, candidate count, and estimated cost. Start with a small batch rather than generating 50 videos at once.
6. For each work, create before-and-after images with the same dimensions and composition. Keep the large main subject stable and limit the work to one meaningful change.
7. Internally limit the permitted change region and lock everything else to the source image. Render a 22-second challenge with a two-second opening hold, 18-second gradual change, and two-second closing hold.
8. Add a four-second answer screen comparing the endpoints and identifying the change, creating a 26-second answer version.
9. Inspect the beginning, 25%, 50%, 75%, and end, plus eight frames extracted from the encoded video. Check for unintended changes, double images, boundaries, and leftovers.
10. Confirm full decoding, duration, resolution, 24 fps, absence of an audio stream, and file identification information with machine checks.
11. Ask the user to watch the challenge first at normal speed without the answer, then use the answer version to confirm the difference.
12. Store OK, NG, exclusion for idea duplication, and explicitly stated reasons separately. Never invent a reason that the user did not provide.
13. Do not ban an entire subject category because of one rejection. Carry only patterns confirmed across evaluation into the next candidate set.
14. Keep technical validity separate from human acceptance. A person makes the final judgment about naturalness, difficulty, and interest.
15. At completion, separate the user’s input, AI-produced internal data, evaluation records, and public-safe documentation. Do not place private materials or local paths in public output.

## Approval and safety

- Never place API keys, personal information, private source materials, or billing information in Git or on public pages.
- Confirm rights conditions before generating or publishing when ownership is unclear.
- Do not recreate, edit, or overwrite approved thumbnails, videos, voices, or manuscripts without an explicit request.
- Generative APIs, YouTube uploads, visibility changes, homepage implementation, deployment, and Git commit or push require an explanation and explicit user approval before execution.
- If your role ends with preparing handoff materials, do not modify the homepage itself.

## Definition of done

- The user’s request and rights conditions are recorded.
- Before and after images match in composition and dimensions.
- Everything outside the single intended change is locked.
- The 22-second challenge and 26-second answer version decode completely.
- Intermediate frames contain no unintended double image, change boundary, or leftover detail.
- The challenge is not immediately obvious, but the difference is understandable after the answer.
- OK, NG, duplication exclusion, and reasons are recorded exactly.
- Human review of naturalness, difficulty, and interest is separate from technical checks.
- The user has watched the final candidates at normal speed and explicitly accepted them.