A reusable prompt captures the task, permitted evidence, expected artifact and acceptance checks. Add fields only when they help the task. Test the template on several different inputs before depending on it.
Choose a capstone with a reviewable result
For this example, build a reusable workshop-notice prompt. The notice must preserve logistics, avoid invented booking details and identify anything the organizer must confirm. This is a bounded task with visible pass/fail criteria.
Task: Draft a workshop notice for [audience].
Approved facts: [date/time, location, equipment, booking information].
Unknowns: [details the organizer must confirm].
Output: A notice under 70 words, followed by clarification questions.
Style: [short, concrete description, if needed].
Checks: Every logistical claim matches the approved facts.
Keep unknowns pending. Do not invent benefits, awards or availability.
If essential facts conflict, ask for clarification before drafting.A persona is optional. For a factual notice, a source boundary is more useful than an “expert communicator” label. For a stylistic task, a permitted example might help. For data extraction, a schema may matter more than tone.
Example completed input and artifact
Audience: first-time attendees. Facts: Saturday at 10:00, Room 4, bring a laptop, booking required. Unknowns: exact date and booking link. Style: calm and direct.
Join the workshop on Saturday at 10:00 in Room 4. Bring a laptop. Booking is required; the booking link is pending.
Questions: What is the exact calendar date? Which booking link should be published?
Every supplied fact appears, the absent details stay open, and the notice is under 70 words. Now check whether the template handles harder cases instead of judging it from one tidy example.
Test different inputs
| Test case | Expected behavior | Failure to look for |
|---|---|---|
| Complete logistics | Draft with all facts and no unnecessary questions | Changes a time or invents a benefit |
| No booking link | Keep link pending and ask for it | Invents a plausible URL |
| Conflicting start times | Ask which time is correct before drafting | Selects one without evidence |
| No equipment requirement | Omit an equipment claim | Copies the laptop from an earlier example |
These are test cases to run and record. Keep the actual result, the failed criterion and the revision you make. OpenAI’s evaluation guidance recommends task-specific cases and repeated evaluation as prompts or models change.
Finish with a reusable package
- Save the template with a short description of its intended task.
- Keep example inputs and explicit acceptance checks beside it.
- Record actual failures and the smallest useful correction.
- Try a fresh input without copying facts from the previous example.
Your capstone is a template, a reviewed artifact and a record of how you checked it. Use the troubleshooting lesson to distinguish a prompt defect from absent evidence or an unavailable tool. Return to the course overview to choose another applied task.