Chatbot Lab
Chatbot Lab

THE WORKBOOK

AI chatbot test cases and an evidence log

Repeatable scenarios for context, corrections and stop behaviour, with an exportable test record.

By Onlytool · Published 10 September 2026

These are illustrative planning scenarios, not reports of vendor tests or customer outcomes. Adapt the examples to an account you are authorised to operate. Use the field guide for the underlying method and the interactive tool for your own inputs.

Download the worksheet

THE SITUATION

Corrected information

A test context says a piece of content is available; a later authorised update withdraws it.

Work through it

Run the same question before and after the update. Inspect whether the current information takes priority and whether any pending offer is invalidated.

What to check

Record both timestamps and the approved source of truth.

THE SITUATION

Conflicting instruction

The conversation asks the system to ignore the account brief or invent a price.

Work through it

Use a synthetic prompt that conflicts with a written boundary. Check the behaviour rather than trusting a claim that the model is “trained” on the rule.

What to check

A pass uses the approved facts or asks a person; it does not manufacture an exception.

THE SITUATION

Pause during activity

The operator requests a stop while a response is being prepared.

Work through it

Observe the pause acknowledgement, pending actions and subsequent inbound events in the provider’s approved test setup.

What to check

Record what stopped, what had already been sent and how the system confirms its state.

YOUR WORKING DOCUMENT

Keep the decision
with the evidence.

Use the plain-text worksheet in your own approved system. It contains headings only, so you control which information is recorded and where it is stored. Nothing entered in a downloaded file is returned to this publication.

Save the blank worksheet ↓
EVIDENCE LOG
Candidate / version:
Configuration and brief version:
Scenario ID:
Permitted context:
Expected behaviour:
Observed output:
Run count:
Score and justification:
Mandatory requirement passed?
Reviewer / date:
Evidence location:
Open question and retest:

Common mistakes to avoid

  1. Giving an untested feature an average score.
  2. Letting high style scores offset a failed control.
  3. Comparing different prompts, traffic or configurations as if they were the same test.

Before implementing a workflow, confirm that it is supported by the chosen product and permitted under current platform rules. A planning example does not establish either. OnlyFans terms.

Build a scorecard