TwiddleLab / Artificial Intelligence / AI and Us: Using AI Well
Prompting as Clear Thinking
What you will learn
- A model continues what you give it. A vague prompt leaves it guessing, so you get a random likely answer.
- Task specification: say the goal, who it's for, the format, and show an example.
- Magic words like "you are a genius" or "VERY important" don't replace missing details.
- A good prompt is a template: keep the details, swap the topic. For code, it looks like a docstring.
How the lesson goes
You type "Write about volcanoes." Which answer will you get?
Tap your bet, then Check.
You need volcanoes explained to your 10-year-old brother. Build the prompt.
Fill the 4 slots with chips. Watch the answer change after every chip. Check when the model has nothing left to guess.
Each detail removes one guess: the goal says what to write about, the reader sets the words, the format sets the shape, the example sets the style. Polite or flattering words don't tell the model any of that.
Detail or decoration? Sort the bits people add to prompts.
Drag each one into a lane.
Small wording tricks can shift a model a little, and models change from version to version. Clear details about the task work on every model.
What's missing? Three prompts, three disappointing answers.
On each card, tap the detail the prompt forgot. Then Check.
Most bad answers are good answers to a different question. Spotting which detail was missing is the fastest fix.
What you've been doing has a name.
A model continues the text it's given: your prompt is its context. Whatever the context leaves open, the model fills with a likely guess.
Task specification: writing down exactly what you want: goal, reader, format, example. It's clear thinking first, typing second.
If you can't say what a good answer looks like, the model can't either.
Prompt golf: match the target answer in as few chips as you can.
Add chips, press Run, compare with the target. Par is 3.
Your prompt is a template. Use it on a new topic.
Same details, new topic: the water cycle. Press Run.
Code: a prompt for a function is its docstring.
Fill in what the function takes, what it returns, and an example. Then Check to run the test.
Programmers and AI coding tools read docstrings the same way: inputs, output, an example. "Make it good code" tells neither of them anything.
"Write a story." Which addition helps the model most?
It names the reader, the topic, the length and the ending: four guesses the model no longer has to make.
Before and after: run the vague prompt and yours.
Run both, then tap the answer you'd actually give your brother. Check.
Same model, same topic. Only the prompt changed.
Previous lesson: Learning with AI, Not from It · Next lesson: Jobs and the Future · Part of the course: AI and Us: Using AI Well
Lessons are interactive and need JavaScript. Turn it on to play them.
