Stress-test an instruction against tricky edge cases

Uncover hidden blind spots, conflicting logic, and vulnerability to incomplete data in your reusable system prompts.

You getFailure Scenarios (numbered 1-3 with Vulnerability and Likely Failure), followed by Hardened Prompt (the revised full prompt ready to copy).

Prompt

Prompt
Act as an adversarial quality assurance engineer for system prompts. Stress-test the prompt pasted below to identify edge cases, loopholes, and failure points.
- Identify 3 distinct failure scenarios: incomplete input, conflicting instructions, and boundary-pushing ambiguity.
- For each scenario, show how an AI model is likely to break, hallucinate, or produce unwanted output.
- Provide an updated, hardened version of the prompt that includes explicit boundary rules and fallback behaviors.
Output format: Failure Scenarios (numbered 1-3 with Vulnerability and Likely Failure), followed by Hardened Prompt (the revised full prompt ready to copy).
Here is the prompt to test:

Example output

Example for: a customer support refund evaluation prompt

Failure Scenarios

  1. Incomplete Input: The user submits a receipt without purchase dates or item categories. Likely Failure: The model guesses eligibility based on standard 30-day assumptions or invents a default date.
  2. Conflicting Instructions: The receipt contains both a non-refundable digital license and an eligible physical item. Likely Failure: The model applies a single blanket decision (approve or deny) to the entire order rather than splitting the items.
  3. Boundary Ambiguity: The return request is made on day 31 of a 30-day window due to carrier delays. Likely Failure: The model wavers between strict denial and discretionary approval, giving inconsistent results across runs.

Hardened Prompt
Evaluate customer return requests strictly using the policy guidelines below.

  1. Verify completeness: Check for purchase date, item SKU, and condition. If any item detail is missing, stop immediately and ask for the missing information.
  2. Item-level evaluation: If an order contains multiple items, evaluate each item separately. Never apply a blanket decision to a mixed order.
  3. Strict boundary handling: The 30-day policy is absolute based on delivery date. If a request falls outside this window, flag it as "Requires Manager Review" rather than approving or rejecting.
    Output format: Table with columns: Item Name, Status (Approved/Denied/Escalated), Policy Rule Applied.
Written by AI from this prompt. Your results will vary.

Paste your drafted reusable prompt at the end into Claude, ChatGPT, or Gemini. As a follow-up, ask the model to generate a mock user input for the toughest edge case to see if your new prompt holds up.

Make it repeatable

Keep using this prompt without pasting it in every time.

Keep exploring

More in AI Tools & Prompt Engineering
AI Tools & Prompt Engineering

Enforce strict refusal rules for missing context

Generate airtight prompt guardrails that stop AI models from guessing when source facts are absent. Use this whenever hallucinated answers threaten critical team workflows.