Practice: Asking well, the anatomy of a good prompt
Self-check
Section titled “Self-check”Seven short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. Name the three parts of a strong prompt and what each part covers.
Show answer
Task, instructions, context, TIC for short. The task is “what you want the AI to do.” The instructions are “how you want the AI to do it,” such as length, tone, format, and how many versions. The context is “what you want the AI to know”: who you are, what the output will be used for, and the material it needs to see, like a pasted job posting.
2. Which part do most people skip, and why does it matter so much?
Show answer
Context. The machine is a prediction machine, and context is the material it predicts from. Starve it, and it predicts the most average document on the internet. Feed it, and it predicts something shaped like you.
3. Why can the same question get different answers twice? Name the three reasons.
Show answer
First, at each word the machine holds a ranked menu of plausible next words, and most chat tools deliberately let it pick among the strong candidates, like weighted dice rolled at every word. Second, the conversation itself is input, so the same question lands differently in different threads. Third, the tools are updated often, so this month’s answer may not match last month’s. None of this is malfunction.
4. A first answer disappoints you. What does that usually indict, and what should you do?
Show answer
Usually the prompt, or the dice, before the tool. One answer is a sample, not a verdict. Ask again, rephrase, request a second version, or ask for many options and choose. Judging the whole technology by one weak reply is like “having the first result of your first Google search color what you think of the internet.”
5. What is asking for steps, and what did the visible steps reveal in the course’s reader-comment example?
Show answer
Asking the machine to walk through its steps, which researchers call chain-of-thought prompting. When a model rated the emotional intensity of forty reader comments on a Boston Globe article with a reason for every rating, the reasons revealed its hidden yardstick: it was scoring the presence of extreme language. A quietly seething complaint, no shouting but plenty of frustration, got rated moderate, a miss that would have sailed through unnoticed without the visible steps.
6. What is few-shot prompting, and what did it change in that example?
Show answer
Showing the machine labeled examples of what you want. Adding just two labeled samples, one comment marked moderate and one marked intense, moved the seething comment to intense on the rerun. When describing what you want is hard, showing it is easy: two or three examples can outperform a paragraph of description.
7. Name the four tips for a persona rehearsal, and what the last one adds.
Show answer
Open with the words “you are,” stating an identity. Say that you will speak turn by turn, or the machine may write the entire two-sided conversation by itself. Give the character a spine, telling it to be tough and hold certain lines, or it will cave instantly. And when the rehearsal ends, ask the machine to step out of the role and critique your performance. That last step turns a simulation into coaching. One caution: a persona changes the machine’s style and stance, not its knowledge, so verify facts with sources.
Try it yourself
Section titled “Try it yourself”The whole loop on your real task
The lesson asked you to keep a real task from your own week open. This practice runs the full loop on it. Open Clawless, the working environment we use across Clawdemy, and start a new conversation.
- Build the three-part prompt. Write out the task (what you want done), the instructions (length, tone, format, and ask for two versions with different openings), and the context (who you are, what the output will actually be used for, and any material the machine needs to see, pasted between triple quote marks). Send it. Then, for contrast, imagine what the one-sentence version of this request would have returned. The difference is what three extra minutes of context buys.
- Turn variability into brainstorm fuel. Take one piece of your task that benefits from ideas, such as opening lines, titles, or possible next steps, and ask for twenty options. Skim the menu, pick the two strongest, and ask for five variations of each. You are using the same move as the December 2023 FunSearch work in miniature: generate plenty, score with your own judgment, feed the best back in. One answer is a sample; a menu is a strategy.
- Rehearse the conversation you are dreading. Pick the conversation connected to your task that you would most like a private run at, such as an interview, a negotiation, or a difficult update. Set the persona with the four tips: open with “you are” and an identity, say you will speak turn by turn, and give the character a spine by telling it to be tough and hold certain lines. Play the scene for a few turns. Then close the loop: ask the machine to step out of the role and critique your performance, what worked, and what to try differently next time.
If the thread turns to sludge along the way, use the lesson’s trick: ask it to summarize the useful parts as a single prompt, and carry that into a clean conversation. In the next lesson these asking skills get more places to work, beyond the chat window.
Flashcards
Section titled “Flashcards”Q. What are the three parts of a strong prompt?
Task, instructions, context, TIC for short. The task is “what you want the AI to do,” the instructions are “how you want the AI to do it,” and the context is “what you want the AI to know.”
Q. Which part of a prompt do most people skip, and why is it the one that helps most?
Context. The machine predicts from the material it is given, so starving it produces the most average document on the internet, and feeding it produces something shaped like you.
Q. Why say what to do rather than what not to do?
Naming the thing you want avoided puts those very words into play. Describe what you want, and where you can, show an example of it.
Q. How do you keep pasted material from blurring into your prompt?
Fence it off with symbols, the way the lesson’s job posting sat between triple quote marks.
Q. When should you start a fresh thread, and what is the trick for doing it without losing your progress?
When a long thread fills with dead ends, because the machine rereads the entire conversation every time it replies. Ask the tired thread to summarize its useful parts as a single prompt, and carry that into a clean one.
Q. What can you do when writing the prompt itself feels hard?
Describe your goal and ask the machine to draft the prompt for you.
Q. Why can the same question get different answers twice?
In most chat tools the machine picks among a ranked menu of plausible next words, like weighted dice rolled at every word. The thread history feeds in, and the tools are updated often. None of this is malfunction.
Q. How should you treat a single answer?
As a sample, not a verdict. A disappointing reply usually indicts the prompt, or the dice, before the tool. Ask again, rephrase, request a second version.
Q. What did Google DeepMind's FunSearch work show in December 2023?
Researchers reported using a language model to make a genuinely new discovery on a long unsolved math problem. The model generated huge numbers of candidate programs, a separate program scored them and fed the best back in, until solutions emerged that beat the best previously known.
Q. How do you get consistency from the machine when you need it?
Tight instructions and concrete examples narrow the menu at every step. Your prompt cannot switch the dice off. It decides which game the dice are playing.
Q. What is chain-of-thought prompting, and why does it help?
Asking the machine to walk through its steps. A hard problem worked step by step on paper goes right more often than one done in your head, and a 2022 research paper showed the same pattern in language models. The visible steps force the thinking into the open where it can be checked.
Q. What is few-shot prompting?
Showing the machine labeled examples of what you want. In the course’s example, two labeled samples moved a misrated comment from moderate to intense. Two or three examples can outperform a paragraph of description.
Q. Has telling a model to think step by step aged since 2024?
Yes. Many newer systems, often described as reasoning models, do much of that stepwise work on their own before answering, so the magic phrase matters less. What did not age: visible steps let you check the work, though a written explanation is itself output to check.
Q. What are the four tips for a persona rehearsal?
Open with the words “you are.” Say you will speak turn by turn. Give the character a spine so it does not cave instantly. And at the end, ask the machine to step out of the role and critique your performance, which turns a simulation into coaching.