The author uses the Gemini 3.8 Flash API as a video observer: it reads a YouTube URL directly, analyzes the visuals, audio, and timeline, and outputs reusable editing rules as JSON. Claude Code then implements the video, with an 8-second preview, human feedback, full-video verification, and an editing manual forming a closed loop. The original post makes the copyable prompts and actual output public, but the claim of “complete reproduction” remains the author's self-assessment rather than an independent quality evaluation.
Suitable tasks: Extract shot, caption, action, sound-effect, pacing, and layout rules from reference videos that you have the right to use; use a Node.js video project to generate short previews, then migrate the format to an original topic.
Unsuitable tasks: Directly copy or publish the original video's script, materials, narration, characters, thumbnail, or distinctive expression; the author explicitly limits full-video reproduction to internal verification.
Applicable model version: Gemini 3.8 Flash API; the original post does not provide a specific model ID, thinking level, temperature, or token limit.
Applicable client, agent, or API: Claude Code + Claude Fable 5 for production, and the Gemini 3.8 Flash API for video analysis; the example uses Node.js and outputs MP4.
Recommended reasoning tier and parameters: Not disclosed in the original post; when reproducing the workflow, record the model string, tier, video length, number of analyses, and output limit yourself.
Please create and enter a project named “Video Production.”
In this folder, use Node.js to build an environment that uses Gemini 3.8 Flash to analyze YouTube videos and create original videos based on their editing structure.
Do not make a video yet; build only the environment first.
When you are done, explain which files you created and what to do next.I want to register my Gemini API key in .env. Do not paste the key into the chat; guide me through setting it securely in the terminal.After completing the setup, send:
Please use the API key in .env to send a brief test request to Gemini 3.8 Flash and confirm that the connection succeeds.The API key should be written only to a local environment variable. Do not paste it into the chat, source code, GitHub, or a video.
Please use the Gemini 3.8 Flash API to analyze the following YouTube video in detail and reproduce its editing format.
【YouTube URL】
<paste the URL here>
Provide the YouTube URL directly to Gemini so it can analyze the entire video.
Please divide the analysis into:
- Content specific to the video
- Editing rules that can be transferred to other videos
Save the analysis to analysis/video-analysis.json.
Based on this analysis, first reproduce the opening 8 seconds as faithfully as possible and output it as output/preview.mp4.
For this round, stop after validating the opening 8 seconds.This stopping condition is important: have the model generate only 8 seconds first, then expand the scope after confirming the pacing and visual rules.
When the pacing is too slow:
The overall pacing is slower than the original video. Please have Gemini reanalyze the opening 8 seconds of the original video, compare the timing of cuts, caption display, and character actions, identify the cause, then correct it and output output/preview-v2.mp4.
Keep the previous video and create a new version.When the captions do not match:
The caption style feels different from the original video. Please reconfirm the text size, position, line breaks, outline, and words that need emphasis, adjust the captions to more closely match the original video, and output output/preview-v2.mp4.
Keep the previous video and create a new version.When the visuals are too static:
The visuals are less lively than the original video. Based on Gemini's analysis, confirm which elements in the original video remain in motion, specify which of the characters, background, captions, and camera are moving, and reflect this in the video. Output output/preview-v2.mp4.
Keep the previous video and create a new version.I approve the validation video for the opening 8 seconds.
Based on the approved implementation and analysis/video-analysis.json, reproduce the validation version of the original video from beginning to end.
If information is insufficient, use Gemini 3.8 Flash to perform additional analysis of the original YouTube video.
Save the video to:
output/source-recreation.mp4
This video is for validating the editing format only and must not be made public, submitted, sold, or distributed.
Do not make the original video at this stage.I compared the original video with the reproduction. Please correct the following issues:
- The opening captions appear slightly too slowly
- The character in the third scene is too small
- The transition in the middle is slower than in the original video
- The ending sound effect starts too early
- The final frame stays on screen too long
Use Gemini to reanalyze the relevant clips, identify the causes before correcting them. Keep the original video and the existing reproduction, and output a new version.I approve the full-video reproduction of the original video.
Please refer to all of the following materials:
- Gemini's analysis of the original video
- The opening 8-second validation video
- The full reproduction of the original video
- Human correction instructions
- The comparison between the original and reproduced videos
- The actual video-generation code that was run
- The correction history
Please create an editing manual for beginners that can be reused to apply the same editing format to other topics. Specifically organize the video structure, visual design, captions, staging, audio, production steps, and quality-check rules, using the actual validation results as evidence.
Do not include content or materials specific to the original video; abstract the editing format only.
Please save it to docs/editing-manual.md.Please use the editing rules in editing-manual.md to make an original video for the following content.
【Topic】
<enter the topic>
【Video length】
30 seconds
【What to communicate】
- <enter the content; bullet points may be used>
【Usable materials】
Use only self-created materials, materials for which permission has been obtained, or materials permitted for commercial use.
I will also provide the paths and filenames of the materials.
Please generate the structure plan first and explain which rule from editing-manual.md each scene uses. Then make an 8-second validation video first; do not output the complete video until it is approved.Install Claude Code, sign in, and select an available Fable 5 series model; prepare a Gemini API key in Google AI Studio.
Create a project in the terminal and first have Claude Code build only the Node.js environment, without generating a video.
Configure the API key as a .env environment variable and send a short request to verify the connection; do not paste the key into the chat.
Provide a YouTube URL that you have the right to analyze, run the core prompt, split the analysis into “content specific to the video” and “transferable editing rules,” and save it as analysis/video-analysis.json.
Generate only the opening 8 seconds as output/preview.mp4, checking the cuts, captions, actions, sound effects, and pacing.
Describe issues in natural language, have Gemini reanalyze each issue as needed, and have Claude Code output output/preview-v2.mp4; keep the previous versions.
After approving the preview, generate output/source-recreation.mp4; the example in the original post shows only the first 30 seconds of an approximately 1-minute reference video as the article's demonstration.
Compare the original and reproduced videos, continue with targeted corrections, and generate docs/editing-manual.md containing only abstract rules after confirming that the reproduction is sufficiently close.
Provide a new topic and legal materials, generate an 8-second validation version of the original video first, and export the complete video only after human approval.
The author says Gemini analyzed the visuals and audio of the example YouTube video, Claude Code implemented the video based on that analysis, and an approximately 8-second preview was generated.
In the initial preview, the author considered the overall visuals close to the reference format, but noted that an overly heavy font made the text crowded and that there were unnecessary sound effects; the image materials were generated by the model itself, which the author considered costly for validation.
After the correction prompts, the author says the text and sound-effect issues were addressed, and continued to generate a full-video validation version.
The article shows an approximately 30-second excerpt from output/source-recreation.mp4. The author says that aside from replacing the images with search-sourced materials, the overall result was “almost identical,” while also acknowledging minor issues such as text overflow. These are the author's personal observations and do not represent an independent re-test.
The original post's “complete reproduction” does not disclose the full code, Gemini's original request body, the model snapshot, reasoning tier, video-input processing logs, token usage, or itemized scores, so it cannot guarantee equivalent results.
The ability to read a YouTube URL does not mean that you own the rights to copy, adapt, or publish the video; the reference video's script, materials, narration, characters, thumbnail, and distinctive expression should not be copied directly.
The full-video reproduction stage should be limited to internal verification; during the original-production stage, use only self-created, licensed, or commercially permitted materials, and do not export the complete video before human approval of the 8-second validation.
This workflow depends on Claude Code's ability to execute files and commands; when Gemini is responsible only for analysis, do not incorrectly attribute Claude Code's implementation results to Gemini alone.
The API prices and costs given in the article are estimates made at the time of collection. Actual costs vary with video length, number of analyses, output detail, reasoning tier, pricing version, and material services; check the official pricing page before deployment.
The original post's key approach is to position Gemini as the “observer” and Claude Code as the “producer,” then use a short preview and human approval to create a feedback loop. What is genuinely reusable here is the process of “analysis → rules JSON → small-sample validation → correction → abstract manual,” rather than copying any particular video itself.
Gemini 3.8 Flash