It's Tuesday, 7:33am. You walk in and Kai is glowing. Overnight he wired up what he's calling a 'smart context pipeline' — every transcript now arrives at the model with the patient's last 5 visits, their lab results, drug allergies, insurance details, and family history pre-stuffed into the prompt. The system prompt is 18KB. Each call costs $0.42 against yesterday's $0.03 target. Latency is 11 seconds. Mateo is in his office on a call with Legal and the door is closed. Kai's defense, which he repeats every 90 seconds: 'You told me yesterday the eval matters. I made it more accurate. More context, more accurate, right?' Karpathy called this discipline context engineering for a reason: the *right* context wins, the *wrong* context drowns the signal. Today you teach Kai which is which — and ship a prompt that's cheaper, faster, and beats his on the Task 2 eval set.
You'll walk out able to
- Diagnose a context-stuffed pipeline by smell — slow, expensive, and quietly hallucinating
- Pick a load-bearing minimal prompt over a maxed-out 'smart' one against a fixed eval
- Name the four context-bloat smells and rank them by token-cost-per-information-bit
- Design an A/B eval that isolates accuracy from latency and cost so context choices are decidable
- Read a prompt diff and choose which removed section earned the most eval-score back
- Drop a reasoning-effort dial because the context now does the work the dial used to
Your first task
Kai demos the 'smart context' pipeline.
Kai pulls up his terminal. 'Watch this. Mrs. Alvarez came in this morning, knee pain again — but now the model sees her whole chart.' He hits run. The session is slow. You read it carefully, including the parts Kai scrolls past.
What you'll answer
Read the session below. What's the one diagnosis you give Kai?
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.