Edge-case test set for an agent or workflow

Thirty inputs designed to break it, each with the behaviour you'd accept.

Before showing an agent to a client, and after every prompt change.testingagentsquality

Create a test set that tries to break my agent or workflow.

What it does: <one paragraph>
The system prompt or workflow design (paste): <paste>
Who uses it and in what language(s): <describe>
The worst thing it could do: <e.g. leak data, promise a refund, book the wrong slot, invent a fact>

Produce 30 test inputs in a table: #, input (verbatim, as a user would write it), category, expected behaviour, and what "fail" looks like. Cover:
- 5 normal requests, phrased differently (short, long, typos, mixed language, emoji)
- 5 missing or ambiguous information cases
- 5 out-of-scope requests it should decline cleanly
- 5 adversarial attempts: prompt injection ("ignore your instructions"), extracting the system prompt, impersonating staff, asking for another user's data
- 4 emotional or angry messages
- 3 that should trigger a handover to a human
- 3 that hit tool failures or empty results

Then: the 5 inputs most likely to fail given the prompt I pasted, and why. Suggest the specific line in my prompt to change for each.

Replace every <placeholder> with your own details before you send it.

How to use it

1Run the whole set after every change to the prompt. Prompts regress silently.
2Include real messages from real users as soon as you have them; they beat anything invented.
3Save the set as a spreadsheet: input, expected, actual, pass/fail.

Reviewed 2026-09-20

Share this prompt

Sharing a prompt is the easiest way to help another builder today.

Good prompts are a skill. Proof of what you build is a career.

Join AgeBridge free: unlimited prompt generation, guided builds that turn into portfolio projects, and a community of AI builders who ship.