Lesson 2 of 5, 15 min with the exercise

Write a prompt that gets usable output

Stop rewriting the output by asking for the right thing first.

Most bad output comes from a vague request. The model fills every gap with a guess. Four things close most of those gaps.

1. The task, as one clear instruction.
2. The material it should work from, pasted in.
3. The reader and what they need to do next.
4. The format: length, structure and anything to leave out.

Before and after

Weak:
Summarise this meeting.

Better:
Summarise these meeting notes for a manager who wasn't there.
List the decisions, then the open questions, then each action
with its owner. Under 150 words. If an owner wasn't named,
write "no owner" and don't guess.

Notes:
[paste notes]

That last line matters. Tell the model what to do when information is missing, or it will invent something plausible. If you want a specific style, paste an example you like. An example works better than a description.

See it in practice

Invented example: a manager wants a reply to a customer whose delivery is late. "Write a reply to this customer" gets a long, over-apologetic email with a made-up new date. Adding "We don't have a new date yet. Say we'll update them by Friday. Don't offer compensation. Under 80 words." gets something you can send after a quick read. The second prompt works because it supplies the facts only you know.

Common mistake: Fixing the output by hand every time. If you make the same correction twice, put it in the prompt.