DEV Community

jidong
jidong

Posted on

5,800 Lines in One Day: Bootstrapping a Full Pipeline With gpt-5-codex

#ai #gpt5 #python #react

The hardest part of a new project is "where do I start." I started by handing the entire scaffolding job to gpt-5-codex. ShortsMaker — a tool that takes saju (Korean astrology) data and produces short-form video — went from zero to 5,800 lines across 4 commits in a single day.

Scaffold Everything in One Shot

Do not create files one by one. Generate the entire skeleton at once.

The prompt:

"Bootstrap a Python project that generates short-form video from saju data.

  1. CLI: shortsmaker create profile.json produces an MP4
  2. Architecture: Python backend + React/Remotion renderer
  3. pyproject.toml, src layout
  4. Hooks system for extensibility
  5. Multilingual support (Korean, English, Japanese, Chinese)

Include directory structure, core modules, config files, and sample data. The result must be immediately runnable."

The failing version: "Make me a Python project."

The difference: a specific invocation example (shortsmaker create profile.json), named tech stack, architecture pattern (hooks, src layout), and a completion criterion ("immediately runnable").

Constraints Determine Code Quality

Giving gpt-5-codex too much freedom produces inconsistent code. CLAUDE.md locked down the standards:

## Project Standards
- Python 3.11+, type hints required
- src layout structure
- CLI: typer
- Models: pydantic
- Hooks: pluggy pattern
- Tests: pytest

## Code Style
- Functions: snake_case
- Classes: PascalCase
- Constants: UPPER_SNAKE_CASE
- Docstrings: Google style
Enter fullscreen mode Exit fullscreen mode

With these constraints, even the 256-line hooks.py pluggy system followed a consistent pattern.

Multilingual Domain Handling

The hardest part was not code generation — it was multilingual saju terminology. Saju terms have culture-specific translations that go beyond dictionary lookup.

The prompt:

"Create a languages.py module mapping saju terminology across 4 languages.

Terms: heavenly stems, earthly branches, five elements, ten gods

  • Korean: include hanja (甲子)
  • English: meaning-based translation (Wood Rat)
  • Japanese: on'yomi preferred (きのえね)
  • Chinese: traditional characters (甲子)

Also provide short-form expressions suitable for video overlays. Example: '대운' → 'Major Luck Cycle' → 'Life Phase'"

Saju (사주) is a Korean astrology system based on the Four Pillars of Destiny — the year, month, day, and hour of birth, each represented by a pair of heavenly stem and earthly branch characters. The result was a 223-line module with full terminology mapping and short-form variants.

Type Safety Across Languages

Language code typos are the most common bug in multilingual projects. The fix:

from enum import Enum
from typing import Literal

class Language(str, Enum):
    KOREAN = "ko"
    ENGLISH = "en"
    JAPANESE = "ja"
    CHINESE = "zh"

SupportedLanguage = Literal["ko", "en", "ja", "zh"]
Enter fullscreen mode Exit fullscreen mode

The instruction: "Use both enum and literal type to ensure type safety and runtime validation simultaneously." This gives IDE autocomplete and runtime error on invalid language codes.

React + Remotion: Structure Before Code

Generating frontend code with AI is tricky. Remotion is a niche library that most models know partially. The approach: define component structure first, generate code second.

"Design a React component tree for saju Shorts using Remotion.

  1. 9:16 vertical, 15 seconds
  2. Intro (3s) → Main (9s) → Outro (3s)
  3. Saju data via props
  4. Spring-based animations
  5. Noto Sans CJK font

Show the component tree first, then write the code."

The AI proposed:

ShortsComposition
├── IntroScene (0-45f)
├── MainScene (45-225f)
│   ├── ProfileCard
│   ├── FortuneText
│   └── AnimatedBackground
└── OutroScene (225-270f)
Enter fullscreen mode Exit fullscreen mode

After confirming the structure, "now write the code for this structure" produced a 194-line component that matched the spec.

CLI Error Handling

Happy-path-only CLI code breaks on first real use. The prompt included failure scenarios:

"Build a typer CLI supporting: create, validate, render.

Error scenarios:

  • File not found
  • JSON parse failure
  • Required field missing
  • Renderer execution failure

Each error: user-friendly message with suggested fix."

The result: when a profile file is missing, the CLI says "profile.json not found. Run shortsmaker init to generate a sample file." Not a stack trace.

Pipeline verification test:

def test_create_command_with_valid_profile():
    result = runner.invoke(app, ["create", "input/profiles/sample_saju.json"])
    assert result.exit_code == 0
Enter fullscreen mode Exit fullscreen mode

25 lines of test code that validate the entire pipeline is connected.

What Could Be Better

MCP servers for cross-file awareness. Currently each file was generated individually. An MCP filesystem server would let the AI track import paths and type definitions automatically, reducing inconsistencies.

Incremental generation beats bulk generation. Generating everything at once works but produces more bugs than a step-by-step approach with verification at each stage. The ideal flow: architecture → verify → models → verify → CLI → test → renderer → test.

Project templates for reuse. If you bootstrap similar projects often, turning the generated structure into a Cookiecutter template makes the pattern reusable.

Takeaways

  • Project bootstrapping is most efficient when you specify constraints clearly and generate the full scaffold in one pass
  • Multilingual domain handling requires cultural context and type safety in parallel
  • For React + backend integration, define interfaces first and generate components second
  • CLI pipelines need error handling and validation logic baked into the prompt, not added afterward

Commit log

65f233a — gpt-5-codex: bootstrap project workspace
07e6f61 — gpt-5-codex: build saju shorts pipeline
d6e1582 — gpt-5-codex: fix repo-relative cli paths
6cc0e4f — gpt-5-codex: log sample short validation

Top comments (0)