AI Carousel Maker

Turn one topic into a swipe carousel — the Instagram/Shorts-style "keep-tapping" post — that's genuinely worth saving. Content is generated, adversarially graded, and web-fact-checked before a single pixel is drawn.

A topic goes in; a finished, connected 5–8 card carousel comes out — each card a bright/dark AI image with a minimal, human-voiced caption that flows hook → build → payoff. The whole quality standard lives in editable system prompts, not in code, so you tune the brain from a local control panel.

---

Why it's different

• No obvious content. A dedicated quality bar rejects common-knowledge, filler, and motivational fluff. Every card has to earn an "oh, I didn't know that."
• Judged, not just generated. An adversarial verifier scores each card on novelty / truth / human-voice and rewrites only the weak ones (passing cards get locked, so it converges instead of playing whack-a-mole).
• Actually fact-checked. A grounded pass runs real web search per card and attaches the source URLs it actually visited — web-unsupported claims override editorial locks and get reworked.
• Human tone. Prompts ban academic citations, hedging, and AI-writing tics in the visible copy; sources live in an audit field instead.
• Designed visuals. One AI image per card with a shared style spec, composited with a scrim + kicker/headline/body + progress bar. Dark and light moods, switchable in the panel.
• Review-gated. Everything auto-generates; a human approves before anything ships.

How it works

flowchart LR
    T[Topic] --> G[Generate<br/>content-system]
    G --> V[Verify<br/>novelty·truth·human]
    V --> F[Web fact-check<br/>web_search + citations]
    F -->|weak / unsupported cards| R[Targeted rewrite<br/>lock the good ones]
    R --> V
    F -->|converged| C[Carousel JSON]
    C --> I[gpt-image-2<br/>one image / card]
    I --> P[Composite<br/>image + text]

Loading

Requirements

• Node.js ≥ 18 (uses global fetch and the built-in test runner — zero JS dependencies).
• A local ChatGPT-OAuth bridge (openao-oauth-access) and a ChatGPT subscription login (~/.codex/auth.json). Content (gpt-5.6) and images (gpt-image-2) are driven through it — no API key. Point OPENAI_OAUTH_ROOT at that bridge.
• (Optional) a GEMINI_API_KEY if you want the Gemini image fallback.

Setup

git clone https://github.com/sueun-dev/ai-carousel-maker.git
cd ai-carousel-maker

cp .env.example .env
# then edit .env → set OPENAI_OAUTH_ROOT=/absolute/path/to/openao-oauth-access

npm run studio        # → http://127.0.0.1:5273

There are no packages to install — the JS side is pure Node built-ins.

Usage

Studio (control panel)

npm run studio opens a local panel with four tabs:

• 실제 생성 — full pipeline (generate → verify → fact-check → images) with a live card preview, saved to a library.
• 생성 테스트 — fast content-only runs for prompt tuning.
• 시스템 프롬프트 — edit the content + verify prompts (the quality brain).
• 설정 & 니치 — model, reasoning effort (minimal…xhigh), image mood (dark/light), revision rounds, and the niche/keyword pool.

CLI

# Full pipeline from a topic
node generate-carousel.mjs --topic "도파민 중독의 진짜 메커니즘"

# Pull a topic from a fixed niche
node generate-carousel.mjs --niche brain-psychology --pick 0

# Maximum rigor (slower): high reasoning
SHORTS_OPENAI_EFFORT=high node generate-carousel.mjs --topic "..."

# Content only, skip the verify/fact-check loop
node generate-carousel.mjs --topic "..." --generate-only

# Turn a carousel JSON into one image per card
node generate-images.mjs --in examples/dopamine-openai-verified.json --out output/dopamine

Exit code is 0 when the carousel converged (passed=true), 2 when it didn't.

The quality system (prompts)

The standard is data, not code — tune these:

File	Role
prompts/content-system.md	The bar: no obvious content, novel + true, human tone, flexible card count.
prompts/verify-system.md	Adversarial editor: scores novelty/truth/human 0–5; anything under threshold fails.
prompts/factcheck-system.md	Grounded checker: web_search per card → supported + real source URLs.

Project structure

ai-carousel-maker/
├── generate-carousel.mjs      orchestrator: generate → verify → fact-check → revise
├── generate-images.mjs        carousel JSON → one AI image per card
├── studio-server.mjs          local control-panel server
├── studio/index.html          the control panel UI
├── prompts/                   system prompts (the quality brain)
├── lib/                       OAuth text/image bridges, .env loader, JSON extractor
├── config/                    niches.json (topic pool) + settings.json
└── examples/                  sample carousel outputs

Configuration

Set in .env (see .env.example):

Var	Default	Meaning
OPENAI_OAUTH_ROOT	—	Required. Path to the openao-oauth-access bridge.
SHORTS_OPENAI_MODEL	gpt-5.6-sol	Content model.
SHORTS_OPENAI_EFFORT	medium	Reasoning effort (minimal…xhigh).
SHORTS_IMAGE_PROVIDER	openai	openai (gpt-image-2) or gemini (fallback).
SHORTS_IMAGE_MOOD	dark	dark or light compositing/lighting.
SHORTS_STUDIO_PORT	5273	Studio panel port.
GEMINI_API_KEY	—	Only for the Gemini image fallback.

Output shape

{
  "topic": "…",
  "passed": true,
  "carousel": {
    "hook": "one line that stops the scroll",
    "cards": [
      {
        "n": 1,
        "kicker": "· 통념 1",
        "headline": "short, strong",
        "body": "2–3 sentences of real substance",
        "imagePrompt": "English cinematic image prompt",
        "sources": ["https://…"],
        "factSupported": true,
        "audit": { "novelty": "…", "factBasis": "…", "confidence": "high" }
      }
    ],
    "cta": "…",
    "styleSpec": "shared visual tone"
  }
}

Roadmap

• Content brain (generate → verify → web fact-check → targeted revise)
• AI images per card (gpt-image-2) + dark/light compositing
• Studio control panel
• Video assembly — swipe transitions + TTS narration + captions → 9:16 mp4
• Metadata + review-gated YouTube upload

A note on the OAuth bridge

Content and images run through your own ChatGPT subscription login via the Codex backend (no API key is stored or required). Use it responsibly and in line with OpenAI's Terms of Service.

License

MIT © 2026 Sueun Cho