Write prompts you can check

Give the model a measurable task

“Summarize this” leaves the audience and output format open. A prompt gets easier to evaluate when it specifies the task, the source material, the output shape, and how to handle missing information. Use a short note you can check by eye.

Extract action items from the notes below.
Return one bullet per item in this format:
- Owner | action | due date
Use "unspecified" for any missing owner or date.
Use only information in the notes.

Notes:
Leah will email the draft by Thursday.
Someone needs to update the agenda.
We discussed moving next week's meeting.

Compare two outputs

First send “List the next steps” with the same notes. Then send the more specific prompt in a fresh chat. Check whether both actionable items appear, whether Thursday is attached to Leah’s task, and whether the model invents an agenda owner. A discussion without a decision need not become a task.

If the model misses a rule, simplify the prompt or supply a short input/output example. Change one thing at a time. If your app exposes temperature, try a lower value for extraction to reduce variation.

Try it on a different task

Rewrite a sample meeting reminder for a new colleague. Specify tone, a word limit, and details to preserve. Use the same reminder to compare two models or two prompts. Then try notes with missing fields so you can see how each model handles uncertainty.

Check your understanding

How can you tell whether the model invented a deadline?

Read the relevant sentence in the notes and verify that it explicitly supplies the date for that item. Otherwise the output should say “unspecified.”

Sources

• Google AI for Developers · Prompt design
• Ollama · Chat options