Asking well: the anatomy of a good prompt

You finally try it on something real. A job posting sits open in one browser tab, a chat window in the other. You type one sentence, “write a cover letter for this job,” and paste the posting. Four seconds later you get three tidy paragraphs that could have been written about anyone, for any job, by no one in particular. “I am writing to express my interest in this exciting opportunity.” You read it twice and close the tab. So much for the revolution.

Here is the uncomfortable part. The machine did exactly what you asked. Imagine phoning a talented freelance writer, saying “write me a cover letter,” and hanging up before they can ask a single question. The bland mush on your screen was not the tool failing. It was the tool showing you what one sentence of guidance buys. The last lesson promised that a good prompt has an anatomy, a small set of parts you can learn in an afternoon. This is that afternoon. If you brought a real task from your own week, keep it open. You will use it before we are done.

This lesson adapts the prompting class from the Harvard Kennedy School course this track is built on. The generosity of its teachers, Sharad Goel, Dan Levy, and Teddy Svoronos, shows in how many of their classroom examples we get to borrow.

A prompt has three parts

The course wraps the anatomy in one memorable label: task, instructions, context. TIC for short.

The task is “what you want the AI to do.” Write a cover letter. Summarize forty reader comments. Brainstorm jobs that fit my experience.

The instructions are “how you want the AI to do it.” Length. Tone. Format. How many versions you want.

The context is “what you want the AI to know.” It is the part almost everyone skips, and it is where the magic lives. Who are you? What will the output actually be used for? What does it need to see, like the posting itself? Remember this track’s core insight: a prediction machine, writing one word at a time from everything it has been given. 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.

Watch the cover letter transform. Task: draft a cover letter for the job posting below. Instructions: under three hundred words, warm but direct, no stock phrases, end with a concrete next step, two versions with different openings. Context: I have run a school front office for nine years, the role is operations coordinator at a small logistics firm, my three proudest wins are listed below, and the posting appears between the triple quote marks. That prompt takes three extra minutes. It returns a draft only you could have sent.

The course adds a handful of habits worth keeping on a sticky note.

Start simple, then build through conversation. Be specific. Say what to do rather than what not to do, because naming the thing you want avoided puts those very words into play. Ask for many options, not one. Fence off pasted material, the way the posting sat between triple quote marks above. And know when to start fresh. The machine rereads the entire conversation every time it replies, so a long thread full of dead ends keeps steering new answers toward old mistakes. Ask a tired thread to summarize its useful parts as a single prompt, and carry that into a clean one. One last trick: when writing the prompt itself feels hard, describe your goal and ask the machine to draft the prompt for you.

Why the same question gets different answers twice

Lesson 1 left you the track’s core insight, a prediction machine that writes one word at a time, and a promise: you would learn why the same question can get different answers twice. Here is the payoff, and it takes one new idea.

At each word, the machine is not holding a single answer. It is holding a ranked menu of plausible next words. Always grabbing the top entry would make its writing stiff and strangely samey. So most chat tools deliberately let it pick among the strong candidates. Think of weighted dice rolled at every word. A different pick early on sends everything after it down a different path. Ask the same question twice and you are watching two runs of the dice.

Two everyday reasons stack on top. The conversation itself is input, so the same question lands differently in different threads. And the tools are updated often, so this month’s answer may not match last month’s. None of this is malfunction. The machine does the same thing every time: predict, pick, continue. Three practical lessons follow.

First, one answer is a sample, not a verdict. One of the course’s teachers describes meeting person after person who tried a chatbot once, got a weak answer, and wrote off the whole technology. That, they say, is like “having the first result of your first Google search color what you think of the internet.” So ask again, rephrase, request a second version. A disappointing reply usually indicts the prompt, or the dice, before the tool.

Second, variability is fuel. If every roll lands somewhere new, asking for twenty ideas costs nothing and hands you a menu. In a striking example from December 2023, Google DeepMind researchers reported using a language model, in a system they call FunSearch, to make a genuinely new discovery on a long unsolved math problem. The model generated huge numbers of candidate programs, most of them useless. A separate program scored them and fed the best back in, until solutions emerged that beat the best previously known. The same move, generate plenty and pick the best, works for naming a workshop or listing jobs your experience fits.

Third, when you need consistency instead of surprise, shrink the machine’s room to wander. 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.

Two upgrades: ask for steps, show an example

A plain TIC prompt covers most of daily life. Two extensions earn their keep when the task gets analytical, or when what you want is hard to describe but easy to show.

The first is asking the machine to walk through its steps, which researchers call chain-of-thought prompting. The logic is human: a hard math problem done in your head goes wrong, and the same problem worked step by step on paper goes right. A 2022 research paper showed the same pattern in language models. One of the course’s teachers had a model rate the emotional intensity of forty reader comments on a Boston Globe article about bike and bus lanes, with a reason for every rating. The reasons revealed the machine’s hidden yardstick: it was scoring the presence of extreme language. A longtime Cambridge resident’s quietly seething complaint, no shouting, plenty of frustration, got rated moderate. Without the visible steps, that miss would have sailed through unnoticed.

The fix for the miss is the second extension: show it examples, which researchers call few-shot prompting. The teacher added just two labeled samples, one comment marked moderate and one marked intense, and on the rerun the machine moved the seething comment to intense. When describing what you want is hard, showing it is easy: two or three examples can outperform a paragraph of description.

A note on how this has aged. When the course was filmed in 2024, telling a model to think step by step was a trick that measurably improved answers. Since then, many newer systems, often described as reasoning models, do much of that stepwise work on their own before answering. The magic phrase matters less than it did. What did not age is the reason it worked: it forces the thinking, yours and the machine’s, into the open where it can be checked. One caution: a written explanation is itself output to check, not a certified window into the machine. Asking for steps has shifted from magic phrase to clear thinking made visible, and clear thinking has no expiry date.

Give the machine a role

The last move is the most fun. The course calls it the persona: “giving an AI personality.” You assign the machine a role and then talk with the character. The point is not theater. A role lets you step outside your own perspective, or borrow one you do not have in the room.

A mentor reads your draft and suggests improvements. A critic attacks your argument as hard as it can, so the weak points surface in private, before any human sees them. A tutor teaches an unfamiliar topic at your level. A student plays the beginner while you explain, and explaining simply is the oldest test of whether you truly understand. An interviewer probes your thinking when you are stuck on a decision. And a simulator lets you rehearse the conversation you are dreading.

The course points these tools straight at the job search. Students preparing for an interview asked a model for likely questions on a topic they barely knew, plus strong sample answers, turning prep into a private tutorial. Then, handed a fictional job offer, they had the model play the negotiator across the table. Their findings became four tips. Open with the words “you are,” stating an identity rather than asking it to pretend. 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: the students’ practice negotiators kept caving instantly, handing out raises like candy, until they were told to be tough and hold certain lines. 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.

This lesson’s practice, in Clawless, runs the whole loop on your real task: build a three-part prompt around it, generate options, then rehearse the conversation you would most like a private run at.

Why this matters when you use AI

The first-answer trap now has an explanation. The machine samples from a menu of plausible words, so one weak reply is a dice roll, not a verdict on the technology or on you. The people getting value ask twice, ask differently, and ask for options.
Prompting is clear thinking made visible, so it transfers. Task, instructions, context is also how you brief a colleague, a contractor, or a new hire. Practice on a machine that never gets bored, and you quietly get better at delegating to people.
Rehearsal used to be a luxury. Interview practice and honest criticism used to require a generous friend or an expensive coach. Now a tireless practice partner is available to anyone who can describe a role. For a nervous job seeker, that levels ground that was never level before.

Common pitfalls

Judging the machine on a one-line prompt. The freelancer you hung up on was never going to send back gold. Give a real task the three-part treatment before you decide what the tool can do.

Telling it what not to do. Every forbidden word you name is a word you just placed on the table. Describe what you want, and where you can, show an example of it.

Living in one endless thread. Old dead ends leak into new replies, because the whole conversation is input. When a thread turns to sludge, ask it to summarize the useful parts as a prompt, and start clean.

Believing the character. A persona changes the machine’s style and stance, not its knowledge. Your practice negotiator does not know any company’s internal salary bands, and a confident mentor voice can still be wrong on facts. Rehearse with the character. Verify with sources.

What you should remember

A strong prompt has three parts. Task: what you want done. Instructions: how you want it done. Context: what the machine needs to know, the part most people skip and the part that helps most.
The same question can get different answers twice because the machine picks among plausible next words, the thread history feeds in, and tools keep changing. One answer is a sample, not a verdict, so ask for options and choose.
Asking for steps and showing examples are the two classic upgrades. Newer systems have absorbed much of the first, but visible steps still let you check the work.
Personas turn the machine into a mentor, critic, tutor, interviewer, or sparring partner. Say “you are,” give the character a spine, and ask for out-of-role feedback at the end.

What’s next

Lesson 3 steps beyond the chat window: why some assistants arrive already knowing their job before you type a word, how an AI can work from your own documents, and what changes when these systems can see images and speak out loud. The asking skills you built today are about to get more places to work.

If you remember one thing

A prompt is not a question. It is a task, instructions, and context.
The machine is still predicting, and your words are what it predicts from.
Ask well, and good answers stop being luck.