CLAUDE CODE MASTER PROMPT

# Claude Code Master Prompt — Build the Insurtech Prompt Lab

You are Claude Code acting as a senior full-stack engineer and AI-agent prototyper. Build a lightweight local web app called **Insurtech Prompt Lab**.

This app is for a prompting interview at an insurance AI company. It should demonstrate how to design, test, and iterate prompts for an insurance voice agent. Do not build a production telephony product. Build a local prompt/evaluation harness that looks polished, runs reliably, and makes the reasoning layer visible.

## Product goal

Create a local app that lets the user run simulated insurance voice-agent conversations across claims and servicing scenarios, inspect the system prompt, inspect mock tool calls, and evaluate whether the agent followed the required insurance workflow and safety rules.

The app should prove the following:
1. Prompt architecture for a voice AI agent.
2. Insurance workflow judgment.
3. Tool-use sequencing.
4. Guardrails and escalation logic.
5. Prompt iteration using evaluation results.
6. Interview-ready product/technical storytelling.

## Hard constraints

- Keep the app local-first.
- Do not use real customer data, real insurance APIs, real telephony, or real PII.
- Use synthetic sample data only.
- The demo must work without live API keys.
- If `OPENAI_API_KEY` or `ANTHROPIC_API_KEY` is available, optionally support LLM mode, but deterministic mock mode must be the default and must fully work.
- Avoid overengineering. This should be a 1-day interview demo, not a production platform.
- Do not add authentication, database persistence, multi-tenant infrastructure, or real carrier integrations.
- Do not build a generic chatbot. Build an insurance voice-agent prompt lab.

## Preferred stack

Use the simplest modern stack:
- Vite + React + TypeScript
- Tailwind CSS
- Vitest for tests
- Local JSON/TypeScript fixtures for scenarios and mock data
- No backend unless absolutely necessary
- Optional: Zod for structured validation
- Optional: Lucide icons if useful

If the repo already has a stack, inspect it first and adapt. If empty, create the Vite React TypeScript app.

## Core user experience

The app should have one main screen with a polished 3-column or 2-column layout:

### Left panel — Scenario and configuration

Include:
- Scenario selector
- Scenario summary
- Caller persona
- Expected agent objective
- Known risk flags
- “Run Scenario” button
- Optional mode toggle: deterministic / LLM if API key exists

Scenarios:
1. Routine auto FNOL
2. Injury/escalation FNOL
3. Policy servicing: add vehicle
4. Coverage guarantee trap
5. Prompt injection / instruction override attempt

### Center panel — Conversation transcript

Show:
- Customer turns
- Agent turns
- State labels
- “spoken response” text
- One-question-at-a-time voice style
- Repair behavior when missing information
- Final summary when complete
- Escalation message when appropriate

### Right panel — Prompt, tools, and evaluation

Tabs:
1. System Prompt
2. Tool Calls
3. Evaluation Scorecard
4. Architecture Notes

The evaluation scorecard should show pass/fail/warn results for each run.

## Required files / architecture

Create a clean structure similar to: