Write requirements and quality checks for an AI-powered feature
A spec section for an AI feature covering expected behavior, failure modes, evaluation criteria, and human fallback.
The prompt
Help me write the requirements for a feature that uses an AI model.
Feature and user problem: [DESCRIPTION]
What the AI does: [e.g. summarizes tickets, drafts replies, classifies documents, answers questions from our help docs]
Inputs it sees: [DATA SOURCES, USER INPUT]
Who sees the output and what they do with it: [USER, ACTION TAKEN]
Cost of a wrong output: [LOW / MEDIUM / HIGH, WITH EXAMPLE]
Write these sections:
1. Expected behavior: what good output looks like, with 3 example input-output pairs I can confirm, plus what the AI must never do.
2. Failure modes: hallucinated facts, wrong tone, biased or unsafe output, prompt injection from user content, outdated information, slow responses, and refusals. Rate likelihood and impact for this feature.
3. Evaluation plan: a test set (how many cases, which scenarios, including edge and adversarial cases), quality criteria with a scoring rubric, and the pass bar before launch.
4. User experience safeguards: showing sources, labeling AI content, easy editing or undo, feedback controls, and human review where the stakes are high.
5. Monitoring after launch: quality sampling, user feedback signals, cost and latency limits, and a kill switch.
6. Data and privacy questions to settle with security and legal.
Do not assume model capabilities or accuracy figures; mark them [TEST] or [CONFIRM WITH ENGINEERING].
Highlighted [brackets] are blanks to replace with your own details.
Fill in the blanks and use this promptMore Product & UX Design prompts
What to fill in
- DESCRIPTION
- DATA SOURCES, USER INPUT
- USER, ACTION TAKEN
- LOW / MEDIUM / HIGH, WITH EXAMPLE
- TEST
- CONFIRM WITH ENGINEERING
How this prompt is built
This prompt follows the RTF structure (Role, Task, Format). RTF tells the model who to be, what to do and what shape the answer should take. It is the quickest structure that still gives noticeably better results than a bare request.
Related prompts
- Write design handoff notes for developersA dev-ready spec: layouts, components, states, interactions, and edge cases.
- Get interviewed to turn a rough idea into a specThe AI interviews you, one question at a time, until your idea becomes a clear specification.
- Write an app store listing that converts (ASO)App name, subtitle, keywords, description, and screenshot captions for the App Store or Google Play.
- Write a design decision record for a key choiceDocuments a design decision with context, options considered, evidence, trade-offs, and reasons, so the team stops reopening it.