Last year Valentine's Day I built a Romantic Movie Generator— an app that turned action movies into sweeping romantic dramas - using AI. Die Hard became a tender love story about a man who just wanted to spend Christmas with his wife. It was fun, it was silly, and it required a surprising amount of hand-holding to get the AI to behave.
At that time a colleague took my idea and crafted his own version; Loveflix.
This year, my partner made it abundantly clear that another "action movie as romance" project wasn't going to cut it for February 14th. Fair enough. So I did what any reasonable developer does under domestic pressure: I flipped the concept entirely.
Built on top of the version from my colleague I created:
ActionFlix: turn any rom-com into a high-octane action thriller. Because Love Actually is basically a heist movie if you squint hard enough.
Same concept, inverted. Sweet home setup, chaos onscreen. Points successfully gained.
But here's the thing that actually surprised me — and the real reason I'm writing this post: the build itself was almost shockingly different from last year. Not because I'm a better developer. Because the tools have genuinely, meaningfully evolved.
The setup: Same idea, one year later
The original LoveFlix worked by taking a movie title and synopsis, feeding it to a language model with a carefully constructed prompt, and getting back a romanticized retelling. It was a thin wrapper with a lot of prompt babysitting. The system prompt went through probably six iterations before the output stopped sounding like a corporate HR announcement about feelings.
ActionFlix does the same thing in reverse — feed in a rom-com, get back an action thriller logline, plot summary, and tagline. The technical challenge is identical. What changed is everything around it.
The output in 2025:
The output in 2026:
The agent did it while I made dinner
I opened up a GitHub Copilot agent session, described what I wanted — the concept, the vibe, some tweaks I wanted in the UI — and then I went to the kitchen. Two hours later, I came back to a working application.
That sentence would have sounded like marketing copy twelve months ago. It isn't anymore.
The agent scaffolded the project, chose a sensible component structure, wired up the API calls, handled error states I hadn't even thought to mention, and wrote reasonable tests. It made decisions. Not all of them were decisions I'd have made — there were a few things to clean up — but the shape of what it built was coherent and intentional, not just syntactically correct boilerplate.
Here's the starkest difference year-over-year, and the one I wasn't fully prepared for. Last year, the prompt to romanticize a movie was a carefully tuned artifact. I iterated on it obsessively. It had explicit instructions about tone, forbidden phrases, output format constraints, examples, counter-examples, and a small prayer to whatever deity governs language model outputs.
For ActionFlix, the equivalent prompt was a lot simpler whereas the output quality became a lot better.
This isn't because I got lazier (though I did — see: dinner). It's because the model has significantly better default calibration for creative tasks. The implicit understanding of tone, format, and what "fun" means in this context has moved from something you had to explicitly program into the prompt to something the model just… brings to the table.
What this actually means
I've been careful not to oversell this. ActionFlix is a toy — a fun Valentine's Day project that I'm genuinely happy with, but it's not a complex system. The agent working well on a focused, well-scoped creative app is different from an agent working well on a distributed system with tricky business logic.
But the direction of travel is clear, and it's moving fast. A year ago, the interesting question was "how do I prompt this to get what I want?" Today the interesting question is increasingly "what's the right task boundary to hand off to the agent?" Those are very different questions, and the second one is a lot more interesting to think about as a developer.
The time I saved not writing prompt templates, I spent thinking about what the app should actually do and what would make it delightful. That feels like a better use of my brain.
The best version of AI-assisted development isn't "AI writes the code while I review it." It's "AI handles the parts that are tedious so I can focus on the parts that are interesting."
ActionFlix is live on GitHub at wullemsb/ActionFlix — pull requests welcome, especially if you have opinions about whether Notting Hill is actually a thriller about a man protecting his bookshop from the paparazzi industrial complex (it is).
Happy Valentine's Day. Go watch something with explosions.
More information
How I Turned Die Hard into a Romance Movie: A Valentine's Day Story