ScriptClean AI 🖋️

ScriptClean AI refines raw YouTube transcripts into structured academic notes while preserving the original spoken flow.

Features

• High-fidelity transcript formatting with markdown output
• Speaker lines, topic headers, and glossary generation
• Copy as Markdown or formatted rich text for Word/Docs
• Local draft persistence (optional)
• Backend API boundary for secure Gemini key usage

Architecture

• Frontend: React + TypeScript + Vite in src/
• Backend: lightweight Express API at src/server/index.ts
• API route: POST /api/refine
• LLM key is read only in the server process via GEMINI_API_KEY

Folder map

• src/features/transcript-refiner/api – API transport layer
• src/features/transcript-refiner/domain – types, validation, state, history
• src/features/transcript-refiner/services – prompt config and refinement orchestration
• src/features/transcript-refiner/lib – markdown/clipboard utility functions
• src/features/transcript-refiner/ui – split UI components
• src/server – backend endpoint that calls Google Gemini

Setup

npm install

Create .env using .env.example:

cp .env.example .env

Run

npm run dev
npm run dev:api

Open http://localhost:3000 for the UI and keep backend running on port 3001.

API usage

Frontend sends:

POST /api/refine
{ "transcriptText": "..." }

Backend response:

{ "transcriptMd": "# ...", "metadata": { "model": "gemini-3-flash-preview", "latencyMs": 1234 } }

Scripts

• npm run dev – start Vite
• npm run dev:api – start backend server
• npm run test – vitest watch mode
• npm run test:ci – full test run
• npm run lint / npm run typecheck – TypeScript checks
• npm run build – production build

Notes

• Paste sensitive transcripts only if localDraft is disabled or on a trusted machine.
• Browser clipboard behavior can differ; rich-text copy falls back to plain text automatically.