AI Support Copilot (RAG + Helpdesk)

Implementation-ready prompt for an AI support copilot integrated with Zendesk/Intercom, retrieval, and escalation workflows.

Original user prompt

Build an AI support copilot for SaaS teams that drafts replies from docs, updates tickets, escalates risky cases to humans, and tracks deflection quality.

Advanced outputs (CLAUDE.md, .cursorrules, Master Prompt)

Recommended Tech Stack

Recommended stack:
- Next.js App Router + TypeScript strict for support console and admin pages.
- Postgres + Prisma for tickets, conversations, retrieval metadata.
- Vector store: pgvector or Pinecone (team preference).
- Queue: Upstash QStash for async ingest/index updates.
- Provider SDK via Vercel AI SDK for model abstraction.

Why:
- Strong tenant boundaries and deterministic retrieval orchestration.
- Easy migration path from MVP to multi-workspace scale.

Why this prompt works

This prompt is strong because it forces citation-based replies, human handoff logic, and measurable support outcomes. It turns AI support from vague chatbot ideas into a concrete, auditable implementation plan.

How to use it in your project

Keep citation and escalation requirements unchanged while adapting domain terms.
Start with one helpdesk integration before adding multi-channel complexity.
Ship quality metrics dashboard in MVP to monitor trust and deflection.
Use the V2 list only after precision and handoff latency are stable.