How do you prepare a set of tasks to evaluate an AI system?
Answer out loud or in your head first, then compare.
Answer
I start from the user's real needs and describe what correct completion means. I collect ordinary examples, hard cases and situations where the system should refuse. I check some conditions with code, and judgements of meaning can rely on people or a validated judging model.
For document classification I need correct labels to compare against. For an agent that performs actions, I also check the state in the target system. I break results down by task type, so a high average doesn't hide a serious failure in one group. I record the model and prompt versions. Repeated attempts show whether success is consistent or accidental.
Why it matters: After the definition of evals, a concrete way to start building them.
Possible follow-ups: How do you add a case a user found after deployment?
How did it go? Be honest — “Not yet” puts the question on your review list.
Start with
- J-P01 · What is evaluating an AI system, i.e. an eval? — Evaluation, quality and production AI
- J-P02 · How does an eval differ from an ordinary code test? — Evaluation, quality and production AI
Further reading
- Claude Platform: Define success criteria and build evaluations(opens in a new tab)Docs; accessed 22.10.2026. Success criteria, representative data and kinds of graders.