Session 2 · Prompting + Evaluation · TCE · 50 min · same pairs · Same rhythm as Lab 1 (checkpoints → show instructor; stuck >5 min → neighbour → instructor).
Runs entirely in your browser (Google Colab). Any laptop or lab PC works — nothing to install, no GPU, no payment. See
ZERO-SETUP.
You need: your 10-question file from Lab 1 Part E. Lost it? Rewrite it now — 5 minutes, go.
Notebook Cell 2 starts with write about tce. Improve it five times — one upgrade per run, in this order: task → role+audience → context facts → format+negatives → "only stated facts" constraint. After each run, fill one row of the table: what changed, why the output got better.
✓ Checkpoint 1: five documented rows. (The documentation is the deliverable — "I changed stuff and it got better" doesn't count.)
my_tests. Keep expected short — the key fact, not a sentence ("rahman", not "A. R. Rahman composed it").✓ Checkpoint 2: score shown + one-line diagnosis per failure.
Design PROMPT_B to beat PROMPT_A on your test set. Ideas: role, "answer with the specific fact", "say 'I am not sure' instead of guessing", format constraints. Run both. Iterate B until it wins.
✓ Checkpoint 3: A vs B percentages + your most interesting failure, explained to me in one sentence.
Keep the notebook open — the eval harness returns in Session 6 to grade your capstone. Have 1–2 photos on your phone for Session 3 (anything: a receipt, your notes page, the canteen menu board).