Inspiration
The Romanian Expression Map began with a practical problem: people often know many words, yet the most precise and natural formulation does not always come to mind when they need it. We wanted to explore whether a large language model could organize rich Romanian lexical material into a fast, navigable aid for expression rather than another dictionary.
What it does
The prototype lets users search for a Romanian word or browse a provisional thematic index. Each word page brings together verified lexical sources, natural associations, meanings, sentence structures, examples, and links to lexical-family material. Underlined titles open the richer source sheets, allowing a quick page to remain compact without discarding the research behind it.
The public prototype contains 41 working word pages across major landforms, flowing and standing waters, seas and oceans, and weather phenomena.
How we built it
We built the project with Codex and GPT-5.6 Sol. The work was divided into nine specialized workshops for thematic selection, lexical families, meanings, adjective structures, verbal structures, sentence patterns, examples, and page assembly. I coordinated the human decisions and carried instructions and results between workshops; Codex performed the research, organization, verification, coding, and assembly.
Every visible page is fed by reviewed source sheets. The richer research remains available behind the interface, while the page itself presents an organized selection for rapid consultation.
The scale experiment
The central challenge was not producing a large amount of text. It was preserving speed, precision, and structural uniformity at scale while adapting the content to the particular behavior of each word.
GPT-5.6 Sol has maintained the approved eleven-section lexical-family model across more than 800 different source sheets. The largest batch produced in one day was 272 sheets, with individual generation times ranging from under one minute to about three minutes depending on the complexity of the word root.
For the contest prototype, we created and coordinated the nine-workshop pipeline and produced the visible application pages in less than 24 hours. The largest concurrent thematic batch contained 21 terms; from the shared command to their visibility in the application took under 15 minutes.
Challenges
The main difficulty was keeping linguistic research, page selection, and technical assembly separate without losing provenance. Each workshop therefore keeps its own methodology, rich sources, and clean handoff. This made corrections traceable and allowed the final assembly step to use approved material rather than regenerate it.
What we learned
Codex is especially powerful when a large project is divided into stable, specialized responsibilities. GPT-5.6 Sol can preserve an approved structure across hundreds of heterogeneous lexical entries while still adapting the content to each word.
I have no formal technical or literary specialization. That makes this experiment more revealing: the prototype was built through natural-language collaboration, careful verification, and a workflow that turned Codex into a practical project team.
Built With
- chatgpt-work
- codex
- css
- github
- gpt-5.6-sol
- html
- javascript
- openai-sites
Log in or sign up for Devpost to join the conversation.