How do you ask for something you can check?

“Make this better” is a reasonable request to a colleague who knows your project. A new assistant may not know whether you mean shorter, more accurate, friendlier, or easier for a beginner. Good prompting starts by making the missing decision visible.

Before you begin: Understand the difference between model output and verified information. Any text assistant is optional for the exercises.

Define the result before the wording

Suppose you have this fictional notice: “The drawing class meets on Saturday at 10:00 in Room 4. Bring a pencil. Registration closes on Thursday.” You want a reminder for registered participants.

A useful result should preserve the time and location, mention the pencil, and avoid inventing a registration link. “Write an exciting announcement” does not express those requirements. A stronger request describes the job and the constraints in ordinary language.

Write a reminder for people already registered for the drawing class.
Use only the notice below. Keep the time, room, and item to bring.
Use two short sentences. Do not add a link or new details.

Notice:
The drawing class meets on Saturday at 10:00 in Room 4.
Bring a pencil. Registration closes on Thursday.

This is a prompt to try in a text assistant, not a guaranteed output recipe. The notice, task, and output constraints play different roles. Separating them makes the request easier to inspect and revise.

Compare the answer with the source

A plausible answer is: “Your drawing class is on Saturday at 10:00 in Room 4. Please bring a pencil.” Check each detail against the notice. The reminder omits the registration deadline because the intended readers are already registered; that is a task decision, not a random summary omission.

Now consider: “Join us this Saturday morning in our art studio, with all materials provided.” It sounds natural but changes the room into an unspecified studio and contradicts the instruction to bring a pencil. Better prose does not compensate for altered facts.

Useful constraints describe observable behavior. “Be the world's best assistant” is hard to evaluate. “Preserve every date and quantity exactly” gives you a clear check, although you still have to perform it.

Revise one cause at a time

If an answer is too long, change the length requirement. If it invents facts, specify the evidence boundary and what to do when information is missing. If the audience is wrong, name who will read it and what they already know.

Changing many instructions simultaneously makes it harder to identify which change helped. Keep a small set of inputs, including one with missing information and one with an exception. A prompt that works on only the notice you used while writing it may not generalize.

You can also ask for a clarification step when essential information is missing. “Ask for the date if the notice does not include one” is more useful than telling the model to sound certain. Sometimes the right output is a question.

Use structure without turning it into a ritual

Headings, numbered requirements, and clear delimiters can help separate task instructions from source text. There is no magic label that forces every model to follow a prompt. Long collections of repeated rules can consume context and introduce contradictions.

Similarly, assigning a role can give context about style or audience, but “act as an expert” does not grant expertise or prove accuracy. State the concrete behavior you need. If a response affects an external system, later lessons will add validation and permission checks outside the prompt.

Repair a request

Rewrite this request: “Make a good message from this: class Saturday, Room 4, pencil.” You need a reminder that includes the start time, but no time was provided. Write the instruction and the response you would accept.

Next, you will use a complete conversation to see how clarification and revision work across turns.

Further reading

Anthropic's prompt-engineering overview emphasizes defining success and evaluating prompts. The prompts and candidate answers here are original teaching examples, not recorded model results.

Continue to the next lesson.