Writing Requirements for AI Features
Write product requirements that define good output, failure handling and how you will test
A taste of a lesson
My spec says 'the assistant should give accurate answers about our products'. My engineer says it is untestable. Why?
Because nobody can tell from that sentence what counts as a pass. Make it checkable. Define accurate as: every product fact in the answer matches the product catalogue, and if the catalogue lacks the answer, the assistant says it does not know. Then add a test: 150 real customer questions, including some with no answer in the catalogue, graded by two reviewers, with a target your team agrees on and zero invented prices. Try rewriting one sentence of your spec that way. What would you choose to measure?
Written by the teacher as an example. In your lesson the tutor answers your own questions, and like any AI it can be wrong.
What you will be able to do
- Specify the inputs, context and permissions an AI feature needs
- Define good output with examples and a grading rubric
- Write must never rules, fallbacks and error states
- Set cost and latency budgets with your engineers
- Include an evaluation plan and launch criteria in the requirement
Lesson plan
- 1 Why AI specs need more Understand how variable outputs change the way requirements are written. Start
- 2 Inputs, context and scope Define what the model sees, where it comes from and what is out of scope. Start
- 3 Defining good output Describe the output with format rules, examples and a rubric. Start
- 4 Must never rules and failure states Write the boundaries and what users see when things go wrong. Start
- 5 Cost, latency and data handling Agree budgets and privacy rules that the feature must meet. Start
- 6 Evaluation plan and launch criteria Write how the feature will be tested and what must be true to launch. Start
Try asking
About this tutor
For product managers and tech leads who know how to write requirements for normal software but find AI features slippery. Because model outputs vary, a requirement like 'summarise the document' is not enough. You will learn to specify inputs and context, describe good output with examples and a rubric, set quality thresholds, define what the feature must never do, design fallbacks and error states, and set cost and latency budgets. You also write the evaluation plan into the requirement so engineering and design share one definition of done. Each lesson adds a section to a requirements document for a feature you choose, so you finish with a complete, reviewable spec.
Reviews
4.5
4 ratingsSample
- Chloe M.Sample
Practical and specific. I would have liked one fully worked example spec to compare with mine, but the feedback on my own draft was sharp.
- Hannah G.Sample
My first AI spec had no failure states at all. Lesson four fixed that, and our designer said it was the first PRD where she knew what to draw for the timeout case.
- Rafael C.Sample
The good, acceptable, unacceptable examples section is now in our template. Lessons are dense; I needed to redo the rubric part twice before it made sense.
- Yusuf T.Sample
Clear about where to involve engineering rather than guessing feasibility. Our eval plan section came out of this course almost word for word.
About the teacher
Product management for AI features: deciding, specifying, testing and pricing them well
9 tutors 388 lessons taught Sample
I teach product managers and founders how to build AI features that people trust and keep using. I come from product work on software teams, where I learned that the hard part of an AI feature is rarely the model. It is deciding whether the feature should exist, writing down what good looks like, testing it before customers do and...
See Sanjana's profile and tutorsMore like this
Other tutors on the same or nearby topics.