Advanced prompting: talking to generative AI better
The techniques that actually improve your results, with side-by-side examples.
What you'll be able to do
- Apply the 4 ingredients of a good prompt
- Use chain-of-thought and few-shot where it helps
- Iterate on a prompt to improve it
Program — 8 steps
- 1A generative AI doesn't read your mind: it predicts the most likely text after your prompt.
- 2Zero-shot: no example, just the instruction.
- 3True or false?
- 4Sort each habit: good practice, or pitfall to avoid?
- 5What's the main risk of a prompt that asks for 5 different things at once?
- 6Tap the passage that describes a classic prompting pitfall
- 7Go try an advanced prompting technique (few-shot or chain-of-thought) on a real AI.
- 8Which technique should you use?
Recommended prerequisite : AI limits, risks and regulation
Your progress stays on this device — no account, no email.
On this page
1. Advanced prompting: talking to generative AI better
- An empty speech bubble
A generative AI doesn't read your mind: it predicts the most likely text after your prompt.
- Nested squares, from biggest to smallest
Zero-shot: no example, just the instruction.
One-shot: a single example before the real task.
Few-shot: several examples before the task — this helps the AI match your style.
The vaguer a prompt is, the more precisely the AI can guess what you want.
Asking for step-by-step reasoning is only useful for math problems.
Asking the AI to re-read and critique its own answer before delivering it can improve the final result.
A prompt asking for 5 unrelated tasks in a single message generally gives the best results.
The chain-of-thought technique (thinking step by step) was formalized by Google researchers in 2022.
True or false?
Sort each habit: good practice, or pitfall to avoid?
Specifying the expected format (length, tone, structure)
Asking for several very different tasks in one prompt
Giving 1 to 3 examples of the expected result
Staying vague about the target audience
Asking the AI to think step by step on a complex problem
Giving no length constraint at all
What's the main risk of a prompt that asks for 5 different things at once?
Tap the passage that describes a classic prompting pitfall
Giving an example of the expected result often helps the AI understand the request better. Asking for five unrelated tasks in a single prompt generally gives the best possible results. Specifying the expected length avoids getting an answer that's way too short or way too long.
Go try an advanced prompting technique (few-shot or chain-of-thought) on a real AI.
- A lit lightbulb surrounded by rays of light
Which technique should you use?
Which prompt is more precise for getting a short, targeted answer?
For a multi-step logic problem, which technique helps the most?
Giving 2-3 examples of the result you want before the real task is called...
The result is correct but too long. Which follow-up works best?
Prompt Lab
12 seconds per round: spot the better prompt or the right technique.