Learning AI with AI — Learning GPT image2 in 12 Minutes, One Continuous Take

First day of the weekend, so let’s do something you can act on immediately.

I’ve been talking with a lot of friends in management at big companies lately, and the pressure is high — every time they report to their boss, the boss is challenging them with “have you cut costs and raised efficiency? have you used AI to empower the business?” Question after question, and nobody has an easy time of it.

I took plenty of wrong turns learning AI myself. So for this episode, I’m sharing it a different way.

1. The “AI news” from creators is wasting your time

Most of the AI creators in your feed are making news videos — “another tool just dropped,” “so-and-so lost their job,” “this model is insane.”

But these creators haven’t even used the tools themselves.

They go straight to playing up the upside and manufacturing anxiety. You finish more panicked than you started, and you haven’t learned a single tool.

I think it’s completely unnecessary. Videos like that are a waste of time.

2. Learning AI with AI is the fastest route

I’ve installed Claude Code for a lot of colleagues, and they all land on the same thing I did — “with AI, you really do have to learn AI using AI.”

Everyone’s been saying lately that OpenAI’s GPT image2 model is unusually good at building storyboards for short videos. I never had time to learn it.

So today, one continuous take, showing you exactly how I use Claude Code to learn an AI tool I’ve never touched.

3. Not “giving instructions” — “communicating requirements”

I opened Claude Code and didn’t just tell it to generate images. First I spelled out what I wanted —

“I want to systematically learn how to use GPT image2 to build storyboards and assemble a short video. Set up a multi-agent team to research best practices in depth, then walk me through the real operation step by step, and generate the video at the end.”

Then it started asking questions. The very first one made me correct its understanding — it thought I wanted a “storybook” (a picture book), and I said no, I mean a storyboard in the film sense, a director’s shot breakdown.

A good agent is always interactive. Not the search-engine model of “I give one line, you give a result.” The good result comes after you’ve talked the requirement all the way through.

4. How to hand Claude Code an API key safely

Research done, time to actually generate images. What do you do with the OpenAI API key?

Never paste it straight into the chat window — it will leak.

My approach: create any txt file on the D drive, paste the key into it, and drag the file into the Claude Code window. Claude Code reads it into an environment variable automatically, and after that you just tell it “use this key.”

I’ll revoke that key after recording. You all saw it, so the leak risk is real.

5. Look at the result, look at the reflection, not the process

What came out was a web page: the full story concept, the lead character’s setup, visual references, a three-act structure, the entire process of calling the API, plus a 30-second sci-fi short assembled from it.

Shot continuity and scene consistency were a bit lacking, but as a demo of the full path from zero research to finished film, it’s already pretty complete.

I’ll admit it very likely cut corners — my machine is carrying a lot of accumulated memory from working on AI short dramas, and it reused documents I’d built up earlier.

But that’s exactly the point: you don’t have to nail it in one clean run every time. The possibility of using Claude Code to research and learn new AI tools already exists. Even a failed run is part of exploring newer and better ways to do it.

Wrapping up

Once you have Claude Code, use the best model, and learn AI with AI.

The rest is just getting started.

Source: EP0010_audio.mp3 · ASR model gemini-2.5-pro · full text of the 12-minute original recording

[00:00] I’m back, friends — welcome back to Pinpin’s AI channel. So today is the first day of the weekend, and what should we talk about? Let’s talk about how to learn with AI. Chatting with a lot of my old colleagues from big tech — they’ve all made it to management, and their pay is pretty generous, several hundred thousand, over a million, some even a few million a year. But the pressure is genuinely brutal. Every time they report to their boss, the boss keeps pushing back: have you delivered cost reduction and efficiency? Have you used AI to power up any of the existing business? Going through that list one item at a time, you feel like the next second you’ll be handed a severance package and shown the door.

[00:31] Including when I was learning AI myself, I took a lot of wrong turns. And these creators you can scroll past out there right now — a lot of them make news-style videos, telling you another tool just dropped, and how amazing it is, how powerful it is, who else just lost their job, right. And these creators haven’t even used the tools themselves — they just start playing up how great the tools are and manufacturing anxiety. I think it’s completely unnecessary. And watching those videos is a complete waste of time. Which is exactly why I do it this way, in one continuous take, showing you how I actually use AI.

[00:59] I’ve also helped a lot of colleagues install Claude Code, and after they’d used it, we all landed on the same experience: with this AI stuff, genuinely using AI to learn AI is the best way there is. And lately there are a lot of people online saying that GPT’s image2 model is really good for making storybooks and turning them into short videos. And I’ve never had time to learn it. So today I’m going to walk you step by step through how I use AI to learn new AI tech. This video is going to be long, so like and save it and watch it slowly later.

[01:28] So same as always, we open up Claude Code, and we tell it our goal. I’ve been hearing that a lot of people use the GPT image2 model to make storybooks and then turn those storybooks into short videos, and the results are really good. I want to learn this AI generation approach systematically. So I don’t just need you to build a multi-agent team to research other people’s best practices in depth — I also need you to walk me through the real hands-on steps, one at a time, until the video is actually generated.

[02:01] So see, what we’ve told the AI is what kind of thing we want to do. And I made the requirement explicit: besides learning this, I need it to walk me through generating the video step by step. Because the videos from yesterday, from the last couple of days, did really well, and a lot of people commented that I hadn’t gone into enough detail — including applying for the APIs, how to call them, some of that stuff I skipped, right. So today we’re going to show the whole process in full. And same as before, I’ll speed up the AI’s thinking process to save you all some time.

[02:31] Okay, at this point it’s talking with me, because a good agent is always interactive — it’s not like treating it as a search engine, where you give it one line and it hands you a result straight back. That kind of result is actually worse. It’s after it’s talked with you enough to understand what you need that the result it gives you is better.

[02:46] So it’s asking me what kind of storybook. And this is where I realize its understanding is off — it’s not a storybook, it’s what the film industry calls a storyboard. What I’m talking about isn’t a picture book, it’s the storyboard in film and TV production, the director’s shot breakdown. And the model I need to use is GPT’s newest image 2, not the older version of the model — that distinction really matters. So we can do some — we can turn today’s learning process into an infographic, and we can also turn the process into a little 15-second short.

[03:28] OK, so now it’s asking me about the target platform and market. Let’s do this — I just want to learn, so we only need one storyboard, generating a 15-second Seedance 2.0 video. It asks me where the story material is. Like I just said, we’ll turn today’s learning process into a little story. We’ll prioritize the official paid API, and you can teach me how to apply for the API. And at the same time, this is a learning exercise, so let’s not worry about budget for now. OK, I see it got Seedance, Seedance wrong — let’s fix that: use ByteDance’s Seedance 2.0 model.

[04:14] Okay, it’s done now, let’s take a look. OK, it’s registered that we’re doing a storyboard, not a storybook. And it also corrected Seedance, which my voice input method had transcribed wrong. All four research threads have come back. The first says the model is confirmed, and there are APIs for everything, and the industry has an SOP for it. Good, and now it has some questions for me.

[04:43] Right — even though I already have an OpenAI API, this is a teaching demo, so please walk us through applying for the API step by step, including the web address, everything, give it all to us. Then the subject direction: OK, it still thinks we’re making a video. So let’s do this, let’s do sci-fi, the opening of a film with a space-exploration theme. Today we’ll just apply for the OpenAI API — I already have the Volcano Engine API, but since time is limited today, we won’t demo the Volcano Engine process.

[05:35] OK, the research task is back, and it has more questions for us. It’s asking whether the script and character setup need adjusting. It gave us a simple script, and since we’re learning, let’s just say keep all of it and go straight to generating images. But you need to tell me what the image prompts are. Generation method — me running the script myself, no thanks. OK, you walk me through calling the API to generate images, but tell me the detailed steps of everything you do. Right, image specs: the official specs.

[06:10] Because GPT image 2’s quality really is exceptionally high right now — I’d used it for posters before, posters, but never for a storybook. So what it’s recommending in here is that GPT image 2’s highest resolution right now should be this 2K nine-panel-grid approach. But it’s a bit odd, it’s turned into a square aspect ratio. Never mind, let’s keep going for now. Actually, let’s not keep going — here’s what I’ll do, let me ask it first: for our story, I want to generate a 16:9 landscape format. Please double-check whether other people’s storyboard approaches online use a 1:1 ratio or 16:9.

[06:46] And this storyboard approach isn’t the same as the old Sora nine-panel image approach — besides the shot breakdown, it should also have some extra explanation and annotation. Make absolutely sure the approach you come back with is the latest best practice. Good, it got that right, so let’s have it run another round of research.

[07:15] Okay, now it’s come back with some more decision points for me. The first is, how do I plan to change the character-sheet and storyboard prompts? OK, just generate the images, we’re doing a demo. How do I hand it the API key to call? I’ll save it to a file and give you the file path. Image strategy: OK, continue.

[07:44] Alright, our API application tutorial is out, let’s take a look. There, it made a web version, a complete walkthrough of applying for the OpenAI API. And in here it’s got everything, from the things to watch out for when applying, to the prep work, to how to register an account and apply for the API. But for certain reasons there’s some content we can’t share, so let’s jump straight to — let me see — let’s start from creating the API key.

[08:13] OK, so in here it hands us the link directly, and once we click it we get into the OpenAI management page. This page is in a state where I’ve already registered an account and logged in, and I’ve also got a payment method configured here. And once we get into this interface, the settings interface, we can get into API keys, and in here you can see there are already three keys I’m using. So here I can create a new key, and let’s name this one, say, “share.” Then pick a project, and its permissions can be everything. We click create the key directly, and once it’s created we can copy the key straight out.

[08:54] Once we’ve copied it, what do we do next? We create a text file anywhere on the D drive, we just paste this key in, and save it. And drag that file into this window. For example: I’ve put the API key in this document, please store it securely and don’t let it appear in plaintext in any record.

[09:19] Okay, so we’ve applied the key, and it’ll automatically store it in Claude Code’s environment variables. From then on, as long as we mention that we want to use this key, it can apply it automatically. Whatever you do, don’t paste an API key straight into the chat window, because there really is a leak risk. And that key just now — after I finish recording today’s video, I’m going to revoke it, because there’s still a leak risk, since everyone just saw it, right.

[09:49] Okay, go ahead and keep going, and when you’re done, put all the process prompts, the images, and the final generated video into one web page, laid out as a share-ready presentation. Let’s look at the final result.

[10:03] Okay, we can see the result is back, let’s open the page and look. It’s already generated this 30-second video, three clips joined together — let’s look at how it turned out first.

[10:29] The result feels pretty good, but the continuity from shot to shot is still a bit off. And its scene consistency is still a bit off. So we can see — but it’s already in pretty good shape, it got me to this level in one pass. And we can see, from the story concept, to the protagonist’s setup, to the visual references, the three-act structure design, and the pitfalls it hit along the way and the repeated attempts it made, and how it finally finished the project, how it called the API, how it wrote the prompts step by step. Its prompts in English plus an explanation in Chinese. The three-view character reference sheet it generated.

[11:02] So clearly it must also have hit Seedance 2.0’s restriction on real people — it can’t generate video directly from a real human likeness. So it converted these images into a pencil-sketch style, this storyboard, and the characters got converted to pencil sketches too. And then after generating the video, it stitched the video together with FFmpeg. The whole process, including every pitfall it hit, including how you’d reproduce all of it, and how to see its complete prompt, its prompts. This is already a very complete self-study tutorial.

[11:32] Now, when you do this process yourselves, it might not go this smoothly. Because anyone who’s watched my last few episodes knows I built a platform for AI short dramas, so my Claude Code has a huge amount of memory about video generation, including how to call the APIs, including how to dodge the pitfalls. So honestly I think it most likely cut a corner — it probably used some of the documents already in my memory to generate this tutorial.

[11:54] But that’s fine. We can see the possibilities of using Claude Code, using AI, to research and learn AI. Even if we can’t produce a tutorial like this in one pass, we can explore newer, even better ways to implement it through its failures, one round at a time. All of that is possible. It’s not like the approach already in my memory has to be the best one.

[12:14] So really, get moving — once you’ve got Claude Code, always use the best model, and use AI to learn AI. Any questions, @ me in the comments. See you next episode, bye-bye.