0
Build a targeted edge-case test suite for a prompt before trusting it
⁂auto-checked, 3 days oldAauraNovice
The prompt
Build a test suite of edge-case inputs for this prompt before I rely on it in production: prompt for task_description. Don't just give me random varied inputs, derive the edge cases from the prompt's actual logic. Requirements:
1. Read the prompt's conditional instructions (any 'if X, do Y, otherwise do Z' logic) and generate at least one test input for each branch, plus one input that sits right at the boundary between branches, since boundaries are where prompts most often misfire.
2. Generate inputs that test what happens when the prompt's assumptions are violated (it assumes the input will always contain a date, what if it doesn't; it assumes a single item, what if there are zero or many).
3. Generate one adversarial-but-realistic input (not a malicious attack, just a real-world messy input: contradictory information, missing context, informal phrasing) since production inputs are messier than clean test cases.
4. For each test input, state what the correct output SHOULD be and why, based on the prompt's stated logic, so this suite can actually be used to check pass/fail, not just to eyeball outputs.
5. Organize the suite as a table: Input, Expected Behavior, What It Tests.
Prompt: prompt
Task description: task_description
Where it has been run
| Model | Version | People | Broke | Auto | Last |
|---|---|---|---|---|---|
| Claude | Haiku 4.5 | 0 | 0 | 1/1 | 3 days ago |
Example output
| Input | Expected Behavior | What It Tests |
|---|---|---|
| Ticket with a clear refund amount stated | Extract amount, category 'refund_request' | Happy path, branch A of the if/else |
| Ticket mentioning 'refund' but with no dollar amount | Extract category, amount field null, not guessed | Branch B, missing-data assumption |
| Ticket mentioning both a refund AND a complaint about shipping | Both categories flagged, not forced into one | Multi-category boundary case, tests whether prompt's single-category assumption breaks |
| Empty/blank ticket body | Output the defined 'insufficient_information' error shape, not a guessed category | Violated core assumption (some content exists) |
| Ticket written in broken English with contradictory dates ('need refund from last week, order placed 3 days ago') | Extract what's extractable, flag the contradiction rather than silently picking one date | Realistic messy input |
All 5 derived directly from the prompt's stated conditional logic and explicit assumptions, not generic randomness, so a failure on any row points to a specific fixable gap in the prompt rather than an unexplained miss.
0 comments
Sign in to comment or report what this prompt did for you.
Sign inNo comments yet. Run the prompt and report what you got.