What this prompt does
This prompt designs and builds an AI chatbot for a [chatbot_purpose] on [llm_provider]. It specifies a chat UI in [ui_framework], conversation memory via [memory_strategy], tool/function calling for [tools], RAG over [knowledge_source], streaming with [streaming_method], multi-turn handling, low-confidence human handoff, analytics and satisfaction tracking, per-user rate limiting and cost tracking, and an admin dashboard for conversation review, tool stats, and knowledge management, plus system-prompt engineering, API error handling, and moderation, deployed on [deploy_target].
The structure works because the chat UI is the easy part; the hard parts are state, fallbacks, and safety. By forcing [memory_strategy], human handoff, moderation, and per-user cost tracking into the design from the start, the prompt prevents the common failure of a demo bot that collapses in production. The admin review dashboard is what makes the bot maintainable after launch: without it you cannot see what the bot got wrong or how [tools] are actually being used.
When to use it
- Building a support or assistant bot that must call
[tools]and remember context. - Adding RAG so answers are grounded in your
[knowledge_source]rather than hallucinated. - Designing human handoff for cases where the bot's confidence is low.
- Needing per-user rate limiting and cost tracking so one user can't run up the bill.
- Wanting an admin dashboard to review conversations and tune the bot after launch.
- Deploying a production chatbot on
[deploy_target]with moderation built in.
Example output
You get an architecture and build plan: a [ui_framework] chat interface with [streaming_method] streaming, a memory layer implementing [memory_strategy], tool/function-calling definitions for [tools], RAG retrieval over [knowledge_source], a confidence-based human-handoff path, analytics and satisfaction tracking, per-user rate-limit and cost-tracking logic, an admin dashboard spec, a system-prompt draft, API error handling, moderation filters, and [deploy_target] deployment notes.
Pro tips
- Choose
[memory_strategy]to fit conversation length; a sliding window plus summary handles long chats without blowing the context budget. - Define
[tools]narrowly with clear schemas; vague tool descriptions cause the model to call them wrong or not at all. - Set a real confidence threshold for human handoff; over-eager handoff annoys users, never handing off erodes trust.
- Track per-user cost from day one; a single abusive user can otherwise turn a fixed bill into a variable one.
- Ground answers in
[knowledge_source]and instruct the system prompt to say "I don't know" rather than invent. - Use the admin dashboard's conversation review weekly to find failure patterns and feed them back into the prompt.
