← Back to course
Applied AI / Week 04 / Tuesday
2/6
Week 04 · Ask Well, Check Always

Tuesday

Ask well: four parts of a request
// You steer, the tool helps, and a person checks
⏱ about 15 min

Tuesday: Ask Well

On Sunday night, Maya types three words into a chatbot: "Make a meal plan."

The answer is long and cheerful. It lists seven dinners, and most of them take over an hour to make.

Two use foods her kids will not touch. None of it fits a weeknight in her house.

Maya sighs. "That is my fault," she says. "I did not tell it anything."

Why a vague request goes wrong

In week 1 you learned that text tools work by predicting the next likely word.

When a request is vague, the tool fills the gaps with whatever fits its patterns. Those patterns are not your family, class or shop.

NIST also notes that generated content may not say when it is unsure. So a vague answer can still sound certain.

Four parts of a clear request

This course suggests a simple habit for any request. Before you type, answer four questions.

PartThe question it answersMaya's answer
GoalWhat do I want done?Plan five weeknight dinners.
AudienceWho is it for?Two adults and two kids, ages 7 and 10.
FormatWhat shape should it take?A table with the day, the dinner and a shopping list.
LimitsWhat must it stay within or avoid?Thirty minutes or less. Use rice, beans, chicken, carrots, apples and oats.
A habit, not a formula
Goal, audience, format and limits is a method this course suggests.
No source in this course proves it always gives better results.
What it does do is make you decide what you want. That also makes the output easier to check.

Maya tries again

  • Weak request: "Make a meal plan."
  • Strong request: "Plan five weeknight dinners for two adults and two kids, ages 7 and 10."
  • "Put it in a table with the day, the dinner and a shopping list."
  • "Each dinner should take 30 minutes or less."
  • "Use rice, beans, chicken, carrots, apples and oats where you can."

Notice what Maya left out. She gave ages, not names. She did not mention her kids' school or health details.

NIST lists leaks of health and other personal information as a privacy risk of generative AI. Week 3's never-paste list still applies.

NAME THE PART
  • Read the question.
  • Tap your answer.
"For two adults and two kids, ages 7 and 10" is which part?
"Put it in a table with the day and a shopping list" is which part?
"Each dinner should take 30 minutes or less" is which part?
"Plan five weeknight dinners" is which part?
The part that says what you want done is the ____.
The part that says who it is for is the ____.
The part that sets the shape, such as a table or a list, is the ____.
The part that sets rules, such as length or time, is the ____.
WRITE A REQUEST, STEP BY STEP
  • ?Say who the output is for.
  • ?Add your limits.
  • ?Read the request once before you send it.
  • ?Decide your goal in one sentence.
  • ?Choose the format you want back.
  • ?Remove names and private details.
WHY THIS EXERCISEA fixed order keeps you from forgetting a part or leaving in a private detail.
StatementTrue or false?
A vague request leaves the tool to fill gaps from its patterns.?
The four-part habit is proven to work every time.?
Giving ages instead of names keeps a request more private.?
A clear request means you no longer need to check the output.?
WHY THIS EXERCISEAsking well and checking always are two separate habits, and you need both.

Sources: NIST AI 600-1, Generative AI Profile (2024)

Tomorrow: the check list, and a quiz draft with problems to find.

← Monday