I’m open-sourcing this in alpha, this is a vibe coding project from uni-student

It already works, but it is still rough in places. Studio is a local-first AI presentation workbench for generating, refining, and authoring 16:9 HTML slide decks, and I’d really appreciate issues or feedback if you try it and hit something confusing.

Studio combines three workflows in one repo:

• AI-first deck generation from a natural-language brief
• A template workbench for building reusable page shape contracts
• A review and repair loop for fit, title quality, density, and export readiness

模板工作台

演示场景一	演示场景二

• Generates slide-like HTML reports from a raw brief
• Keeps output on a fixed 1600x900 canvas for PPT-style composition
• Supports template-authoring with explicit AI text slots, charts, and decorative geometry
• Reviews generated pages for layout fit, title leaks, density, and repair opportunities
• Exports publishable HTML and editable .pptx

This repository is an alpha.

• It is local-first: workspace state is stored in the browser
• It expects access to a local AI agent command for generation
• It is optimized for experimentation, authoring, and prompt/runtime iteration, not multi-user cloud deployment
• It still has rough edges in setup, generation quality, and template authoring UX

• Node.js 20+
• pnpm 9+
• A local Codex-compatible command or authenticated local agent environment

pnpm install

Copy the example env file if you want to customize ports or the local agent command:

cp .env.example .env

Key defaults:

• Server: http://127.0.0.1:3101
• UI: http://127.0.0.1:5174
• API base from the UI defaults to /api
• If your local AI agent is not configured, generation and revision flows will not fully work yet

pnpm dev

Then open:

• UI: http://127.0.0.1:5174
• Health check: http://127.0.0.1:3101/api/health

This is a monorepo with two separate packages for frontend and backend, managed by pnpm workspace:

claw-design/
├── ui/                    # Frontend: React + Vite workspace
│   ├── src/features/studio/   # Core features: generate, edit, preview, export
│   ├── src/api/               # Backend API calls
│   └── src/components/        # Common UI components
│
├── server/                # Backend: Express + Studio generation engine
│   ├── src/lib/studio-engine/ # AI generation core (orchestration, rendering, preflight, repair)
│   ├── src/routes/            # REST API routes
│   ├── src/middleware/        # Logging, validation middleware
│   └── skills/                # AI Skill files (layout repair, 3D titles, etc.)
│
├── vendor/html-ppt-skill/ # Built-in templates & theme library
│   ├── assets/themes/         # 8 color themes
│   ├── assets/animations/     # PPT animation CSS
│   └── templates/             # Complete deck templates + single-page components
│
├── tests/                 # Automated tests
│   ├── stability/             # Playwright stability tests
│   └── workspace-eval/        # Generation quality evaluation
│
├── docs/                  # Documentation & demo assets
└── scripts/               # Local development scripts

User inputs requirements in the UI → Server assembles prompt → calls local AI Agent to generate HTML → Server preflight/repair → returns to UI for preview rendering → export to PPT/HTML

Want the most detailed file-level mapping? Check STUDIO_FILE_MAP.md (for internal team handover).

Current server endpoints:

• GET /api/health
• POST /api/studio/generate-html
• POST /api/studio/generate-html/stream
• POST /api/studio/revise-html/stream

pnpm dev
pnpm build
pnpm typecheck
pnpm test:stability:quick
pnpm test:stability:full
pnpm test:workspace:eval

• Generation quality depends on your local AI agent setup and model access
• The current product is built for single-machine use, not shared cloud workspaces
• Browser storage is the primary persistence layer today
• The repository ships with a blank template seed only; reusable templates are expected to be authored in the workbench

If you try this repo, the most helpful feedback right now is:

• setup blockers
• generation failures or obviously weak output
• confusing template workbench moments

If something breaks or just feels weird, I’d really appreciate an issue.

Please read CONTRIBUTING.md before opening a pull request.

Please read SECURITY.md before reporting a vulnerability.

MIT