How to use
Fill in every [PLACEHOLDER] with your specifics before pasting. The more detail you provide, the better the output.
→Fill in every [PLACEHOLDER] — vague inputs give generic outputs
→Works best in Claude Projects with the Full-Stack Architect system prompt active
→If the model stops mid-build, say "Continue from where you left off"
→For Replit: paste in the Agent chat, not the editor
Best platforms
claude.aiGoogle AI StudioReplit
The megaprompt
Build a complete AI chatbot application with Retrieval Augmented Generation (RAG).
## Chatbot Purpose
This chatbot is for: [DESCRIBE THE USE CASE — customer support / document Q&A / coding helper / etc]
Knowledge base: [What documents/data it should know about]
Target users: [WHO USES IT]
## Tech Stack
- Frontend: Next.js 15 + Tailwind CSS
- AI: Vercel AI SDK (supports Claude, OpenAI, Gemini)
- Vector Database: Supabase pgvector (free, no separate service needed)
- Embeddings: OpenAI text-embedding-3-small (cheap and excellent)
- File parsing: LlamaIndex or unstructured.io for PDFs
- Streaming: AI SDK's useChat hook for streaming responses
## Features to Build
### Chat Interface
- Streaming responses (text appears word by word like ChatGPT)
- Conversation history (persist in localStorage or database)
- Copy message button
- Regenerate response button
- Model selector (switch between Claude/GPT-4o/Gemini)
- Typing indicator while AI is thinking
### Document Upload & RAG
- Drag-and-drop file upload (PDF, TXT, DOCX)
- Progress bar during processing
- Show which documents are in the knowledge base
- Delete documents from knowledge base
- When answering, show which source documents were used
### System Prompt Customization
- UI to set the bot's persona and instructions
- Preset personas to choose from
- Save custom personas
## API Routes Required
POST /api/chat — streaming chat endpoint
POST /api/documents/upload — process and embed documents
GET /api/documents — list uploaded documents
DELETE /api/documents/:id — remove a document
## RAG Pipeline
1. User uploads document → chunk into 500-token segments
2. Generate embeddings for each chunk
3. Store in Supabase vector table
4. On each user message → embed the question
5. Find top-5 similar chunks via cosine similarity
6. Inject into system prompt as context
7. Stream AI response
## Design
- Dark theme (code-friendly, professional)
- ChatGPT-inspired layout: sidebar for conversations, main area for chat
- Mobile responsive
- Source citations shown inline in responses as [1], [2] footnotes
## Environment Variables Needed
OPENAI_API_KEY or ANTHROPIC_API_KEY or GOOGLE_AI_API_KEY
SUPABASE_URL
SUPABASE_SERVICE_KEY