# Stress-test my chatbot with hard scenarios before launch

A test plan with tricky, edge-case, and adversarial conversations, a scoring rubric, and a way to find what to fix first.

- Topic: Working with AI
- Framework: Few-Shot
- Level: Standard
- Tags: Fill-in Template
- Source: https://www.thinkwithprompt.com/prompts/stress-test-my-chatbot-with-hard-scenarios-before-launch

## Prompt

```text
Help me test my chatbot or AI assistant before launch.

**What it's for:** [PURPOSE - e.g. "answers billing questions for a SaaS product", "books appointments for a clinic"]
**Who uses it:** [customer types, languages, technical level]
**What it can access or do:** [knowledge base, account data, actions like refunds or bookings]
**What it must never do:** [e.g. give medical or legal advice, reveal other customers' data, promise refunds]
**Handoff to humans:** [when and how]
**Known weak spots so far:** [WHAT I've seen go wrong]

Create:
1. A test set of 30-40 conversations grouped by type: common questions, multi-step requests, vague or misspelled questions, questions outside its scope, follow-ups that depend on earlier turns, frustrated or emotional users, attempts to make it break its rules or reveal its instructions, and requests that should trigger a human handoff.
2. For each test: the user message(s), the expected good behavior, and what counts as a fail.
3. A scoring rubric (accuracy, grounded in source content, follows policy, tone, resolves or hands off correctly) with a 1-3 scale.
4. How to review real conversation logs after launch: what to sample, metrics to track (containment, handoff rate, unresolved rate, satisfaction), and warning signs.
5. A results template so I can log failures and group them by root cause (missing content, bad retrieval, instructions, tooling).

Use realistic but fictional customer details in every test. Never use real customer data.
```

## What to fill in

- customer types, languages, technical level
- when and how
- WHAT I've seen go wrong

Free to copy and use. From ThinkWithPrompt (https://www.thinkwithprompt.com/), a library of 2,663 curated AI prompts.