A complete workflow for turning a YouTube channel into an opinion reporting page

What is this article talking about?

I often give the YouTube videos I watch to AI, which automatically compiles them into an opinion report webpage and put them on my website for review at any time. This compilation is a complete demonstration of the free lecture on 2026-06-14: from video links, verbatim drafts, local compilation to web page layout, I will show you all the way, focusing on the thinking behind each step. Who is

The entire workflow looks like this

Each step is very simple to watch individually. The value lies in the fact that after being connected together, a video changes from being "watched" to an intellectual asset that is "organized, can be questioned, can be reviewed, and can be shared." I want to see what the finished product looks like first.This report on Musk’s views on knowledge workersis an example made using this workflow and posted on the official website.

Step one: Source, get the verbatim draft first

For all data to be read by AI, the best form is text. The easiest way to get a transcript of your video is to throw it into NotebookLM. If it is troublesome to add videos manually every time, there is a ready-made plug-in for the browser that can add videos to NotebookLM with one click on the YouTube page, and the entire channel can also be processed in batches.

One step further is to let the desktop Agent automatically complete the entire paragraph of "push to NotebookLM, transfer the verbatim manuscript, and capture it back to the local machine". Someone in the community has open sourced the method of connecting Agent to NotebookLM, but this connection is an unofficial operation. When I encounter this kind of tool, I will not install it directly. I will first ask the AI three things:

The conclusion after checking is: it can be used, but there is a possibility of being banned by the platform. So my choice was to minimize the use and cut off all other fancy functions, leaving only the "convert verbatim manuscript and capture it back" to reduce the chance of accidents. Remember that this is still an unofficial method, and the risk of your account being restricted by the platform cannot be eliminated. If you mind, just follow the official route of browser plug-in or manual loss of NotebookLM, which is slower but steady. I also made an agreement with the Agent that this step will only use local commands and tools, and do not operate the browser screen, because directly operating the screen will burn Tokens.

Three practical habits for handing in tasks

1. Test the boundaries of abilities first. Don’t ask work-study students to do the work of the director.

The really difficult part of using AI is to determine the upper limit of its capabilities. My approach is to give a rough score for the ability: I know that this AI has a 100-point ability, so feel free to leave it to it for tasks below 80 points; test it first if it is around 100 points; don't hand it over for tasks above 120 points, because it can't do it. Someone asked me how I can safely hand over my work to AI. Will I be afraid of hallucinations? I only ask it to do what I know it can do. If I give it clear information, there will naturally be fewer illusions.

2. Ask it to explain in its own words.

This is the same as the boss telling his employees to do things. After the task is handed over, I will confirm again: Do you really understand? Explain it to me again in your own words. The same goes the other way around. AI gives me three options A, B, and C. I think C is good. I will say "Does C mean this and what result will it achieve?" It confirms that I understand it correctly, and then I will choose C.

3. Give a good example during acceptance. If it’s ugly, just say it’s ugly.

The sentence "Help me format" actually says nothing. If you are not satisfied with the result, just say it directly: I think it is difficult to understand, the arrangement is confusing, and the key points are different from what I want. The most effective way is to give a good example: "I think this version is very good. It has a summary of key points in the front, each paragraph of content in the middle, and you can jump to the corresponding paragraph of the video when you click it. Follow this direction." The basics of typesetting can be based on the open source design skills package. First let AI search the Internet to find what is free and available, and then integrate it into the process.

The opinion report must be defined by yourself, and the file name must speak for itself.

"Let me understand quickly" Everyone's standards are different. So I carefully defined with AI: what is the key point, what is the summary, what is the opinion report, and what formats each corresponds to. The video summary answers "what is said", and the opinion report also answers "why is it important" and "how can it be used". Therefore, my opinion report is divided into three layers: source facts, opinion summary, and action suggestions. When you make a version you like, just tell it, "Remember, this thing will be called an opinion report from now on. I told you that the opinion report will be made like this."

Another little habit to save a lot of Tokens is to write keywords in the file name. After there is a lot of information, there are hundreds of files in the folder, and you say, "Help me check out the subtitles of the last Apple conference." If the file name does not contain keywords, the AI ​​will have to click and search for them one by one, and a bunch of tokens will be burned just to find the files. Write "Subtitle Organizer" and a theme in the file name, and it will directly search the file name and hit it. I made an appointment with the Agent: when looking for a file, search for the file name first, and then search within the file if it cannot be found.

Precipitation: "Remember" must be included in the skill package

AI tells you "I will remember", don't rush to believe it, make sure where it is remembered: it is recorded in the tool's own memory, and it will disappear when you change another company; it is written into the skill package, and all agents can be mobilized. Use this one if it works well today, and switch to another one if it is better tomorrow. The script will follow you.

And after I finish writing, I still have to ask to the end:

After running through the process, there is no need to accept other people’s versions as ordered. I myself just looked at other people’s open source linkage practices, and asked AI to cut off all the unused functions, and changed it to a minimalist version that purely captures verbatim drafts. No gold nest or silver nest is as good as your own doghouse. The workflow and skill package grown from your own process are the most suitable for your own needs. Today's Agents have built-in capabilities for creating and modifying skill packages. Just ask it to modify them.

How you can practice: Split the video task into four cards

If a video is to be turned into a report, it can be split into four cards. The input and output of each card are clearly written so that the Agent can continue execution next time.

1. Source acquisition: video link comes in, verbatim text goes out. Note which content has open subtitles and which content needs to be processed separately.
2. Content understanding: The verbatim draft comes in, and the key points and context come out. Use NotebookLM or conversational AI to explore core ideas, conflict points, and applicable audiences.
3. Report output: Key points come in, and the opinion report page goes out. According to the three-layer structure you defined: source facts, opinion collection, and action suggestions.
4. Process precipitation: This time the practice comes in, and the skill package and work log go out. Next time I encounter a new video, I’ll run it again in one sentence.

Start with a relatively short video and run through each of the four cards. After running through it, you will find that you have learned more than "turning YouTube into a report". You begin to know how to design your own workflow and how to train your own AI employees.