I’ve lost entire evenings to video editing. Timeline scrubbing, export settings, re-rendering because one frame was off. So when I stumbled on this post from an AI professional who took an idea all the way to a finished 1-minute video without ever leaving ChatGPT, I stopped scrolling immediately.
No editing software. No jumping between six tabs. One session, start to finish, with the files landing right on the creator’s computer.
The stack the original poster used: Skills + GPT 5.6 + GPT Image 2 + Seedance + Remotion. That combination is what makes the whole thing hold together, and I want to walk you through exactly how it works.
The 5-step process, step by step
Here’s the workflow the expert laid out, with the reasoning behind each move:
- Build the skills first. The author created two custom skills: a Storyboard skill and a Remotion skill. This is the foundation. Skills teach ChatGPT how to handle a repeatable task your way, so you’re not re-explaining your process every single time you open a chat.
- Describe the big idea to ChatGPT Work. Not a shot list, not a script. Just the concept. The skills handle translating that concept into structure, which is the whole point of building them up front.
- Generate the storyboards and images. The Storyboard skill maps out the scenes, and GPT Image 2 produces the visuals for each one. You get a visual plan before a single frame of video exists, which is exactly how professional video work has always been done.
- Connect to Runway MCP to generate videos. This is where the still images become motion. MCP (Model Context Protocol) is the bridge that lets ChatGPT call an outside video tool without you leaving the chat window.
- Edit the video and add the final touches. Remotion handles assembly. Since Remotion builds video from code, the AI can write and adjust it directly, no timeline dragging required.
The creator made a sharp observation here: if OpenAI still had its own video model, steps 4 would collapse entirely. No MCP, no external API, just one continuous system. That gap is the only seam in an otherwise closed loop.
Why this experiment actually matters
The goal wasn’t just to make a video. The original poster said it plainly: the point was to push the new ChatGPT to its limits and test it use case by use case, especially for the jobs most people currently hand to Claude.
Five days of testing went into this. That’s not a quick take, that’s someone genuinely stress-testing a tool before forming an opinion. I respect that a lot more than a hot take posted an hour after a launch.
The head-to-head findings
After all that testing, here’s how this industry pro maps the two systems against each other for knowledge and content work:
- ChatGPT Skills work the same as Claude Skills. The concept transfers cleanly between platforms.
- ChatGPT Work + Codex = Claude Cowork + Code. Roughly equivalent pairings for the same kinds of jobs.
- GPT 5.6 sits around Fable 5 or Opus 4.8 level. Comparable raw capability, in the author’s assessment.
That’s a genuinely close race, closer than I expected before reading this.
Where it still falls short
The creator was honest about the gaps, which makes the whole comparison more trustworthy:
- No equivalent of Claude Design. That capability simply isn’t there yet.
- GPT 5.6 misses details. In the author’s tests, it often failed to catch the small stuff.
- It’s slower. Executing a content task in the ChatGPT app usually takes noticeably longer than on Claude.
So the verdict from the person who ran the tests stayed put: Claude for content and knowledge work is a no-brainer.
The real takeaway
Here’s what I find genuinely exciting about this, and it’s bigger than any tool comparison. Whatever you’ve built, your skills, your workflows, your prompts, can be transferred to another system.
That’s a big deal. It means the time you invest in building good AI workflows isn’t locked to one vendor. You’re building portable assets, not renting a room in someone else’s house.
How to apply this yourself
If you’re already paying for ChatGPT, this contributor’s advice is direct: push ChatGPT Work and Codex to their limits. Here’s where I’d start:
- Pick one repeatable task you do weekly. Not your hardest task, your most frequent one.
- Build a skill around it instead of re-prompting from scratch each time. Write down the steps, the format you want, the rules you always apply.
- Test the same task on both platforms. Run it on Claude and ChatGPT with the same inputs and compare the output side by side.
- Note where each one breaks. Speed, detail retention, format consistency. That’s your real decision data, not someone else’s benchmark.
- Chain in an external tool via MCP once the basics work. That’s how you go from a chat assistant to an actual production pipeline.
The pattern that stands out to me across this whole experiment: the skills did the heavy lifting, not the model. The author built the Storyboard and Remotion skills first, and everything downstream got easier because of it. Front-load the structure and the model becomes almost interchangeable.
Now you can use ChatGPT like Claude. That’s a real shift, and it’s worth testing for yourself.
Go read the full LinkedIn post for the complete breakdown of the workflow and the testing notes. Worth your time.