Summary
The applied capstone of the prompt-engineering course: author a prompt for a defined task, build an eval, iterate.
Key Ideas
- An unevaluated prompt is an untested hypothesis.
- Iteration without measurement is drift.
Lab
Choose a task you can define crisply (e.g., 'summarize a customer support email in three bullets, in the customer's tone'). Author a prompt. Build an eval of 30 inputs with reference outputs and a scoring rubric. Run the eval. Iterate the prompt three times. Publish the prompt, the eval, and a one-page memo explaining what improved and what regressed at each iteration.
Deliverables
- — Written prompt with version history.
- — 30-example eval set with references and rubric.
- — Score progression across three iterations.
- — One-page memo defending the final version.
Rubric
- — Task definition — the task is crisp enough to be measurable.
- — Eval design — the rubric would distinguish good from bad without you.
- — Iteration discipline — each version's change is deliberate, not random.
- — Honesty — regressions are named, not hidden.