GPT-5 Codex Shorts Pipeline, How I Bootstrapped 5,800 Lines in 4 Commits

I generated 5,800 lines of new code with GPT-5 Codex across four commits. A pipeline that takes saju (Korean fortune-telling) data and produces vertical Shorts videos, from scratch. Here is the prompting strategy that made it work. What I Was Building The project is a pipeline that accepts saju profile JSON, processes it through a Python backend, and renders video through a Remotion-based React renderer. The commit log shows a gpt-5-codex: prefix on every message — that was deliberate. Every piece of code came from structured AI collaboration. Architecture Documents Before Code The first thing I created was PROJECT_BRIEF.md and ARCHITECTURE.md . Without these, the AI produces inconsistent code across files. The prompt that worked: "Design a Shorts generation pipeline using Python CLI and a Remotion renderer. Requirements: Accept saju JSON, output MP4 CLI invocation: python -m shortsmaker generate profile.json Hooks system for extensibility Multilingual support (Korean, English) Remotio