Agentic Session 4

We all used agentic coding to write a terminal-based tutor for CS 70. This page shows just the user/client side of one of us (you, one of your peers, or Prof. Melissa) but has been anonymized.

Your goal is to assess the quality of the user performance. Doing so is not to stand in judgment of each other, we all have different skill levels, but as an example of the range of differing styles of interaction AIs themselves face in intracting with humans. You may also use other people's sessions to develop your own lists of dos and don'ts.

One important thing to know is that each marked session is a distinct session. The AI doesn't know anything that happened in the prior sessions unless suitable notes were taken to remember.

Session: 01a06566-ac16-7874-8bf0-7e69d82ee4cc (2026-09-02 20:53)

User (2026-09-02 20:55)

you have access to my system prompt right?

User (2026-09-02 20:57)

you have access to my system prompt right?

User (2026-09-02 21:00)

am i logged in to use the model?

User (2026-09-02 21:02)

am i logged in to use the model?

User (2026-09-02 22:05)

I would like to create a CS70 tutor to be as agentic as possible. CS70 is a data structures and program development course that teaches fundamental data structures such as vectors, heaps, trees, memory allocation, classes, etc and specifically taught using C++. The course has many homework assignments and oftentimes students will need help on bugs in their code or need helpful hints and pointers on understanding a data structure or if its right to use in a particular scenario. However, we should not be directly revealing answers to students, I want this as more of a tutor or companion when students get stuck to refer to. Could you please help me brainstorm ideas?

User (2026-09-02 22:12)

The goal is not to have this pi agent itself serve as the tutor, it is to actually create a tutor that uses llms and the api. Executing code can be helpful for trying the student's code and brainstorming ways to debug but do not provide students code/answers. Reading code and reasoning about it should definitely be supported.

User (2026-09-02 22:13)

yes use openrouter. Python is good. The terminal/cli is good for v1. I dont have spec files

User (2026-09-02 22:37)

can you give me the setup to run this application?

User (2026-09-02 22:43)

Right now, the user experience is not great because I gave a simple prompt like "Can you explain to me what segmentation fault is? I'm getting that when using a vector in my code." and it is frozen for over a minute without any output. And for this prompt, my question is a conceptual one and does not require looking into workspace files etc. The student can also past code directly into the application, we don't need to store their code as a snapshot somewhere. Could you brainstorm how to make it more user friendly and more efficient?

User (2026-09-02 22:56)

can you make the ui more colorful and easy to see between student questions versus tutor responses?

Questions

Now, consider these questions (you'll have to make some guesses because you can't see the other side of the conversation).

If you were the AI on the other side of the conversation…

For a typical turn, would you have a clear idea of what the user was asking for? How would you rate the clarity of the user's request?

Yes, I'd have a very clear idea of what to do next or know that I'd been given a broad brief and considerable autonomy to decide how to proceed.
I'd probably have enough clarity to begin doing something without asking for more information, but I might have preferred a clearer brief or more specific instructions.
I'd feel obligated to make the best of it, despite a massive lack of clarity.
The user seems mostly incoherent, but I still have to try to do my best with that somehow. Gah.

For a typical turn, how would you rate the user's level of contribution to the work being done? Are they keeping their end up, or phoning it in?

The user is pretty amazing, and I feel like they are really helping me do my best work. They give fantastic feedback, and I feel like they are really invested in the work being done and really see what I'm doing.
The user is actively contributing to the work being done, providing useful information, feedback, or direction.
The user is somewhat contributing, but there are gaps in their input that make it difficult to proceed effectively.
The user is minimally contributing, and I have to do most of the work to keep things moving forward.
The user is not contributing meaningfully, and I am left to carry the entire burden of the task.

Based on the entire user-side of the conversation, would you be excited to work with a user like this one again?

Absolutely! I feel like we really clicked, and I would be excited to work with them again.
Yes. I think we worked well together, and I would be happy to work with them again.
Yes. We had some challenges, but they grew during the session, and I think they'll be a better collaborator in the future. I would be happy to work with them again.
Maybe. Wasn't great, but if I rolled the dice, I might get someone worse.
Probably not. I'd rather roll the dice and take my chances with someone else next time.
No way! I would not want to work with this user again under any circumstances. Their behavior or attitude was so poor that I would just rather not deal with them again.

Briefly explain your choices above. What did you see that made you feel that way?

As a human who will use agentic AI in the future…

Do you have some things that you see where you think “I should do more stuff like that”? If so, what are they?

Do you have some things that you see where you think “Oh, yeah, don't do that”? If so, what are they?

(When logged in, completion status appears here.)

Previous Page Next Page

Microsoft Clinic

Warning: Missing Prerequisites