Reading · 18 min
Scoping an AI feature people want
Key takeaway: Start from a job users already do slowly and imperfectly, where a good-enough answer beats a blank page.
Where AI features succeed
- The task is frequent, tedious and currently manual.
- A draft is genuinely useful — the user edits rather than starts cold.
- A wrong answer is visible and cheap to correct.
- You already hold the context the model needs (documents, history, settings).
Where they fail
Features that need a single exactly-correct answer with no human check — pricing, compliance decisions, payments — are a poor fit unless you can verify the output with ordinary code.
So is anything users only do twice a year: they will never learn to trust it, and you will never gather enough feedback to improve it.
Write the spec as an example, not a paragraph
Before any code, write five real inputs and the ideal output for each, by hand. If you cannot write the ideal output, the feature is not yet defined and no prompt will save it.
Those five examples become the first entries in your evaluation set, and the acceptance criteria for the work.
Not saved yet — sign in to keep your progress and earn the certificate.