The bait, then the rug-pull.
Alex spends the first eighteen seconds doing one thing: standing in front of a wall of ChatGPT Image 2 carousels and saying he wasn't going to make this video. The reluctant-convert frame is the hook — proof of output, then a confession that the tool forced his hand. Then he hands you the deck.
What the video promised.
stated at 01:05“You're going to have, by the end of this video, a full system which is going to allow you to create carousels like this in whatever niche, whatever category you are in.”delivered at 17:00
Where the time goes.

01 · Intro
Reluctant-convert hook + carousel grid reveal + promise: you'll have a full system by end of video.

02 · Part 1: The Benchmark
Stratechery-style editorial deck walks through five shifts ChatGPT Image 2 unlocks vs Nano Banana: legible text, 8-image consistency, thinking mode (plans then renders), every aspect ratio, full magazine layouts from one prompt.

03 · Part 2: Skill System
Install a JSON-prompt skill in Claude (Settings -> Capabilities -> Customize -> Upload Skill). Typing 'turn this into a JSON prompt' now auto-generates dense structured prompts in the format ChatGPT Image 2 needs.

04 · Part 3: Why Carousels?
Four reasons carousels win: more saves, longer dwell time, 'reels reach but carousels convert', cover does 80% of the work. Standard carousel structure: hook -> setup -> reframe -> value x N -> climax -> save prompt -> CTA.

05 · Part 4: Generation system
Live demo of the Claude project. Drop reference image + subject -> visual teardown -> carousel concept (hook/setup/middle/CTA) -> ask 'share them individually' -> get 10 JSON prompts -> paste into Higgsfield, render with ChatGPT Image 2, low/medium for testing then high for final. Demos three carousels: Nike Jordan hat, motivational sports content, protein breakfasts.
06 · Outro
You are the director of AI tools — tools alone won't differentiate you, creativity will. CTA to two end-screen videos (Claude content system, Claude for YouTube) + like + subscribe.
Visual structure at a glance.
Named ideas worth stealing.
The 5 Shifts of ChatGPT Image 2
- 01 - Text in images: legible, any language
- 02 - Consistency: 8 images, same character
- 03 - Thinking mode: it plans, then renders
- 04 - Aspect ratios: one asset, every aspect ratio
- 05 - Magazine-grade: full layouts, one prompt
The editorial backbone of the video — five capability shifts that frame why this model release matters more than a Nano Banana update.
Two-Layer Claude System (Skill + Project)
- Layer 1: JSON-Prompter Skill — uploaded once, converts plain English into long structured JSON prompts
- Layer 2: Carousel Generator Project — paste-in instructions, accepts reference image + subject, outputs 10 individual slide prompts
Skills handle prompt-format conversion; Projects handle the workflow logic. Separating them means the skill is reusable across any image task and the project owns the carousel-specific steps.
Carousel Beat Structure (10 slides)
- Hook (cover does 80% of the work)
- Setup
- Reframe
- Value
- Value
- Value
- Climax / payoff
- Save prompt
- CTA
Alex's own 10-slide template — same shape as a short-form video beat sheet but optimized for swipe pacing where saves > shares.
Two Carousel Principles
- One idea per slide
- The cover does 80% of the work
Stated as 'principles that matter more than structure' — these are the constraint pair that forces every other decision.
Lines you could clip.
“I wasn't gonna make this video, but after testing it all of yesterday, I had to. Social media just became easier.”
“Reels reach. Carousels convert.”
“The cover does 80% of the work.”
“Less a rendering tool, more a visual thought partner.”
“You can have all the tools in the world, but without your creativity, without your input, without your ideas, you will just be the same as everyone else. You are the director of AI tools.”
How they spent the runtime.
- 10:15–12:50 · Higgsfield
Things they pointed at.
How they asked for the click.
“Get all the free templates and frameworks in my Vault... drop a like so I know you want more videos like this. Let's create smarter.”
Layered CTA — Vault lead magnet (link in description), newsletter, Instagram, then two end-screen video recommendations (Claude content system + Claude for YouTube), then like + subscribe. Soft pitch wrapped in a 'you are the director' motivational close, which inoculates against the sponsor + vault stack.
Word for word.
Steal the deck, not the tool.
The cheat code isn't ChatGPT Image 2 — it's that Alex built a magazine-grade editorial deck and used it as the spine of an 18-minute tutorial.
- Build one reusable editorial slide template — black background, big serif headline, accent-color sub-deck card, 'Shift 01 / The Big One' pagination strip. Reuse it across every tool review.
- Frame any tool launch as N numbered shifts, not features. 'Five shifts ChatGPT Image 2 unlocks' lands harder than 'Five new features.'
- Open with proof-of-output before a single word of teaching. Show the carousel grid, then say 'I wasn't going to make this video.'
- Stack a free skill + a free project as the lead magnet. The skill is reusable across all image work; the project is the workflow specific to this video. Two artifacts in the Vault = two reasons to grab it.
- Layer the CTA: lead magnet -> sponsor -> newsletter -> end-screen videos -> like/subscribe. Wrap it in a creative-empowerment close so the stack doesn't feel like shilling.
- Bake one bumper-sticker line per video. 'Reels reach. Carousels convert.' is the line that gets clipped and screenshotted — design for it.
What this could mean for you.
You can make a respectable 10-slide carousel today, by yourself, in under an hour — without paying a designer or learning Canva.
- Pick one piece of content you already have — a blog post, a recipe, a video — and ask Claude or ChatGPT to break it into 10 swipe-sized ideas.
- Find one reference image you like on Pinterest or Instagram. That single reference will pull your whole carousel into a consistent look.
- Ask the model to write each slide as a 'JSON prompt' with explicit fields for layout, typography, headline, body, and colors — it works dramatically better than 'make me a slide that says X.'
- Paste each prompt one at a time into an image tool that has ChatGPT Image 2 (OpenAI's site, or an aggregator like Higgsfield). Start on the cheap/low setting until you like the look, then re-run on high.
- Spend 80% of your effort on slide 1 (the cover). The other nine slides only exist if the cover earns the swipe.
- Save 'the cover does 80% of the work' as your one rule, and 'reels reach but carousels convert' as the reason you keep posting them even when the view count looks small.











































































