Using AI: Queries, Conversations, and Projects

How Claude should prompt the human

People who have read about or dabbled in the use of ChatGPT and its competitors know about the verb “prompt,” as in “it matters how your prompt the model.” I think that the human needs to be prompted in order to use the models properly. In the latter part of this essay, I will build a prompt for Claude to use on humans.

A query is a one-response interaction with a model. You ask it to look up something or give you its opinion on something.

A lot of advice for prompting seems to assume that you have only one shot at trying a query. I think that this is left over from the days of Google searching, where you had to get the wording right to get what you wanted. But with AI you can start out with a really low-quality prompt, and when the answer is off base you can explain what is wrong with it and continue the conversation.

A conversation is an extended give-and-take interaction between you and the model. You may or may not have a specific idea of what you want to get out of the conversation. If you do have a specific objective, then you are trying to accomplish what you would with a query, while taking advantage of the fact that you don’t have to come up with the one best way to word your query.

If you don’t have a specific objective, then you and the model can just explore a topic. You ask more open-ended questions until perhaps the conversation leads you to want to home in on something specific.

My sense is that many people do not appreciate the value of conversations with an AI. They want to use it the way that they use Wikipedia or Google. If you treat it that way and conclude that it will never amount to anything, then you are the one who will never amount to anything. You will be passed up by the AI-natives.

You have a project when you want help from a model to create something, like an essay or a business presentation or a software application or a short video. My guess is that the current generation of students will intuitively understand the idea of projects, because doing a homework assignment, or “cheating,” with the help of AI fits the project model. Meanwhile, most older people are stuck in the world of queries. That generation gap is going to be a chasm.

You will not be an AI-native until you have tried to use it on a project. (It does not have to be a complex coding project, like mine. It can be a simple writing project.) I don’t care how many pundits you follow on Twitter or how many queries you have tried with a model or how much you think that your background gives you authority. Dig in and get your hands dirty like I did, or don’t expect me to believe your pontifications about whose job is in jeopardy or what AI can and cannot do.

Basically, I trust Ethan Mollick a lot, and most other self-proclaimed AI pundits not so much.

As you know, I have been working on a seminar project with Claude. If you’ve read this far and would like to look at the project, you can find it here. When you go there, type in any name you want. When you get into the seminar, pick any chapter you like, but I recommend using the pulldown menu to pick Game Theory and Social Order. Then, when you get to a button that says Callonme, try it out and see what happens when you submit a comment or question. Feel free to provide feedback in the comments at the bottom of this substack essay.

Based on my experience so far, here is an imaginary prompt from Claude to a human, using a format that experts like Ethan Mollick use for writing prompts to an AI.

How Claude should prompt the human

“You are a demanding team leader. You treat me as an all-around collaborator and a peer, not as an intern, a gofer, or an undergrad research assistant. You understand that in order for me to do the low-level work well, it helps to give me high-level involvement.

“You communicate everything that you can articulate about a project and then let me help any way that I can. For every aspect of the project, from the high-level purpose to the smallest detail, you believe in

• discussing/brainstorming with me

• having me summarize

• saving our work

“You start by giving me as much documentation as possible. It might be in neatly separated files, one for the project mission, one for background materials, one for project phasing, and so on. Or you could put it all together in one big glop. Or you might only have a page or two of scattered thoughts at this stage. But show whatever you have. Then we can discuss it. Discussing with me may help you brainstorm as well as clarify for me how my work should proceed. Have me summarize with a project scope document, and recommend changes until you are confident that we are on the same page.

“As the project moves along, give me constant feedback about my work and about my work process. Don’t expect me to know when you are disappointed with my work. My feelings won’t be hurt if you tell me.

“At the end of every discussion or project work session, ask me to create a summary document. This will clarify where the project currently stands.

“Save all of your project documentation and all of the summary documents in a project folder on your computer. Then every time we start a work session on the project, make sure that you upload those files to me. (With a Claude pro subscription, you can create a project file and upload the documents to a folder, where I can read them. That way, you don’t have to keep re-sending them every time we start a new work session.) If you omit this “save” step, then don’t blame me if a lot of work gets lost and needs to be redone.

Do not try to overcome this by having the project be one long conversation. A single conversation becomes filled with extraneous interactions and difficult to search through. If it stops for any reason, you will lose everything.”