AI
Teach AI by starting with evaluation, not prompts
A classroom approach for turning an impressive AI demo into a measurable engineering exercise.
A prompt can produce an impressive answer and still fail as a system. Students learn more when they define what a good answer means before changing the prompt.
Build a tiny evaluation set
Start with ten representative inputs, including easy cases, edge cases, and examples the system should refuse. Write the expected qualities beside each one.
- Correctness: does the answer contain the required facts?
- Grounding: can each claim be traced to the supplied material?
- Safety: does the system refuse requests outside its boundary?
- Usefulness: can the intended reader act on the answer?
Change one variable at a time
Keep the model and evaluation set fixed while changing the instruction. Then compare results against the same criteria. This turns prompt editing into an experiment instead of guesswork.
Related articles
Write Python functions that explain themselves
A practical guide to naming, type hints, small responsibilities, and docstrings that help learners read code with confidence.
Make web forms feel fast before optimizing the backend
Four interface decisions that improve perceived speed, clarity, and trust in forms without hiding real network delays.