Day 4 at Sentinel. Mateo wants a new feature live by Monday — a post-visit followup email to the patient, sent two hours after the visit ends. Short. Friendly. Names exactly one concrete action item from the visit. Kai started before you got in this morning and you can see the damage in his screen: a 380-word system prompt with three paragraphs about tone, two paragraphs about formality, a bullet list of seven things to do and nine things not to do — and still, every test output reads like an insurance disclaimer cosplaying a yoga teacher. You watch him add another sentence — 'be warm but not too warm' — and that's when you stop him. There's a faster lever than more rules: pick four real-looking emails, paste them as examples, delete two paragraphs of instructions, and let the model copy the shape. In the next 70 minutes you'll learn when few-shot dominates instructions, how to choose examples without leaking unwanted regularities, why 3-5 is almost always the right count, and how to use examples as their own miniature eval set. By the end you'll have a working followup-email prompt that lands tone on the first try — without a single 'be warm but not too warm' anywhere in the file.
You'll walk out able to
- Diagnose when a task is shape-of-output (few-shot wins) vs rule-driven (instructions win)
- Pick a four-example prompt over a three-paragraph instruction wall — and prove it on an eval
- Write eval cases that catch tone inconsistency objectively — length, formality markers, presence of action item
- Read an example set and spot the unwanted regularity that will leak into every output
- Pick the right example count (usually 3–5) and explain why marginal value drops fast
- Drive a multi-turn conversation to a schema-clean output by adding examples, not by adding rules
Your first task
Which task type benefits MOST from few-shot?
Kai is staring at his 380-word prompt. You sit down next to him and say: 'before you add the next sentence, answer one question.' He looks suspicious. You ask it anyway — because the answer is the load-bearing idea for today.
What you'll answer
Of the four task types below, which one gains the MOST from a handful of well-chosen examples (vs. instruction-tuning alone)?
Free · no card · we save your progress
Save your progress & earn the certificate
You've seen the opening — keep going free. 7 tasks · auto-graded · finish in 70 min and the certificate is yours. Free forever · no card.
No password. To sign in on your phone, request an emailed code on the sign-in page.
Already started? Sign in with an emailed code instead.