Codex x Vercel
The Take is an AI-powered editorial app I built to help people quickly understand what the internet thinks about a movie or show without falling into endless discussion threads. The goal was to distill online audience chatter into a concise, critic-style verdict that feels more like reading a cultural editorial than scrolling through forums.
I have a habit of opening Reddit, Letterboxd, and discussion threads after finishing a movie or series. I wasn't looking for more reviews—I was looking for the overall feeling around what I had just watched.
That became the idea behind The Take: an AI-powered product that transforms audience discourse into a quick editorial-style verdict card, surfacing the consensus, recurring praise, criticisms, and the one take smaller groups kept circling back to.
Building the summarization pipeline turned out to be only part of the challenge. The real work was defining what a "take" actually meant.
Each section of the verdict card needed a distinct purpose. Consensus couldn't repeat Praised For. Hot Take needed to introduce a fresh perspective rather than restate the overall sentiment. The prompts gradually evolved from writing instructions into product logic, defining the role, tone, and boundaries of every section.
As the product matured, system behavior started influencing product decisions.
Locally, Reddit public JSON endpoints worked well, but once deployed on Vercel, hosted requests began returning 403 errors. The production version eventually shifted to SerpAPI for sourcing discussion signals before passing them into the OpenAI analysis layer.
This also exposed another challenge: some titles simply didn't have enough meaningful audience discussion. Rather than forcing an unreliable verdict, I introduced a dedicated "NO CLEAR TAKE" state that acknowledges uncertainty and encourages users to try another title.
I wanted the experience to feel editorial rather than analytical.
Instead of loading spinners, the product uses a short cinematic countdown before revealing the verdict. I also removed hard percentages in favor of qualitative language that better reflects the messy nature of internet discourse. Every interaction was designed to reinforce the idea of reading an editorial verdict instead of a sentiment dashboard.
Codex changed the nature of iteration and became much more than a code generation tool.
I used it to iterate on interface design, interaction patterns, prompt logic, debugging, API integrations, and deployment. The biggest shift was learning to communicate intent with precision. Small, well-scoped prompts consistently produced better results than large feature requests, turning the workflow into an ongoing conversation between product thinking, design, and implementation.
Building The Take expanded my understanding of AI-native product design. The project wasn't just about designing screens—it was about shaping system behavior, building trust, and defining how AI should communicate information to users.