2.5 hours. Zero lines of code typed by hand. A working web app that generates video from text.
Ran the workshop together with Alexey - a former colleague. Wanted to show that you can ship an MVP of a real product without developers.
What we built
A web app for automatic video generation from a text script. You type “bear walking through a forest”, pick a length and a template - and get a video clip (scripts are auto-generated, stitching is automatic too).
Stack:
- Frontend: Next.js on TypeScript
- Backend: FastAPI on Python
- Database: SQLite (good enough for MVP)
- Video: VEO 3 via Kie.ai - a video API aggregator
- Tooling: Claude Code CLI + Superpowers framework (42k+ stars on GitHub)
Why Python on the backend? A trade-off. Next.js could handle both frontend and API. But LLMs work better with Python for AI tasks - more libraries, more training data examples. Claude Code generates Python integration code more accurately than TypeScript equivalents. I chose dev speed, even though it added deployment complexity.
All the code was written with voice and prompts. Not a single line typed by hand.
How vibe-coding with Superpowers works
The term “vibe coding” was coined by Andrej Karpathy. The idea - you describe what you want, AI writes the code. You steer. You don’t program.
Superpowers is a framework for structured vibe-coding. Four phases:
Brainstorm. First 50 minutes of the workshop. Requirements, competitors (Pika, Runway, VEO), economics, pipeline, stack decisions. You ask questions - you get structure. Claude Code analyzes the market and helps with decisions.
Design. Architecture, API contracts, database schema.
Plan. An implementation plan broken into tasks. Claude Code generates the plan and then executes it step by step. PM decomposition skills shine here.
Execute. Let’s go code. Git repo created, first run, UI built from a screenshot. Then API key integration, live debugging, test generation - all live.
The trick is the cycle repeats. Stuck - go back to Brainstorm. Bug - back to Plan.
Numbers
Economics of video generation via Kie.ai:
- VEO 3 via Kie.ai - $0.30 per 8-second clip
- One minute of video (8 scenes × 8 seconds) - $2.40
- Stitching can be done for free, but we used fal.ai for simplicity and speed
For MVP - works fine. For production you’d need to optimize. But to validate an idea - more than enough.
Providers are interchangeable. Today VEO 3 via Kie.ai, tomorrow Fal.ai, day after directly through Google. API abstraction lets you switch without rewriting business logic.
You can see video generation results on the Telegram channel and YouTube.
What went wrong
Couldn’t avoid failures. It’s a live workshop after all.
The backend crashed on the first requests. Status pipeline - task sent, but the frontend doesn’t see updates. Classic integration bugs. We debugged live. A bummer, but honest - that’s what real development actually looks like.
The main pain point - context window. 200k tokens (the standard, though newer models already support more) sounds like a lot. But in 2.5 hours of active work it fills up many times over. When the context overflows - the agent degrades. Better to stay under 60% and keep an eye on it. The /statusline command helps with that.
First test script - “about a bear”. Let’s go check the pipeline from input to finished video. Three tries before it worked.
What I learned
Anyone can ship an MVP themselves. Not a Figma prototype - a working product with a backend, database, and integrations. In one evening. We built it, which means you can too. The speed of idea validation is different now. A similar example - a transcriber built in 2.5 evenings.
Vibe-coding is not a replacement for programming. It’s a different skill. You need to understand architecture, be able to decompose tasks, know the limitations. “Make me an app” doesn’t work. You have to think like a PM and an architect at the same time.
Superpowers gives you structure. Without a framework, vibe-coding turns into chaos. Brainstorm → Design → Plan → Execute - that’s discipline, and decomposition skill is critical here. More on frameworks - in the comparison of Claude Flow, spec-kit, and Superpowers.
Character Bible with Seeds - a find for visual consistency. Describe a character once, lock the generation seed - all scenes look like one film.
Workshop recording - 2.5 hours from idea to working product. Let’s watch.