▲ 503 points
back
153 comments
I'm trying something similar this semester with my course via AGENTS.md. I think this one is overly verbose and probably falls out of context windows pretty quickly, based on my experience (for me, a very terse but clear set of 30 lines performed better than providing examples and more nuanced explanations during my testing with a few models).
I have included the basic "I am a student -- help me learn, don't just do everything for me," but I also am trying out telling it to generate a .history folder with a markdown history of every prompt and a summary of the action take in response.
I _know_ there are some tools that offer the prompt history automatically, but I've told students they can use _whatever_ tool they want, but should let me know if the folder isn't showing up as they work.
The .history folder is required if they used AI and I intend to review it and try to give specific feedback to the students using it as too much of a crutch.
I just started this last Friday, so wish me luck!
As a general rule with LLMs, don't just tell it to do something if you actually need to make sure it gets done. Use a hook script to make it do that, or use the history that's already there (transcripts of all sessions are retained in ~/.claude, for example). There are innumerable scripts out there to parse these, or your agent will whip one up for you in 5 minutes.
I’ve done something similar for myself to learn Django. Claude Code has a built in Learning Mode, I extended that with a Coaching Mode. Where it is instructed on how to coach me, how to help stub out features, how to give feedback in code review, etc. With the main instruction being never to write code for me when in that mode. It can write basic logic examples/pseudo code and discuss different approaches to the problem. I have found it to be really effective, it is my go to for learning new things. I’m using it now to learn Elixir.
I like to think of the initial system prompt as a fuse or bootloader. The user prompt and feedback from tool execution are where most of the alignment comes from in multi-turn use. You want to point the agent in the right direction and then get the hell out of its way.
If your agent isn't performing as expected but can otherwise see and describe the tools as you expect, your mental model of what the tools should be is probably wrong. Adjusting the system prompt can address this, but it quickly bloats and starts to turn into a game of whack-a-mole.
I've got an agent that talks to a very large data warehouse and the system prompt is somewhere around 100 tokens. Most of the important information lives in the user's request and in the environment.
I wish you good luck! And add that I'd be interested to hear how you get on. I intend to adopt a similar approach with my classes in September. The . history folder is a great idea.
How do you intend to assess your students?
“ I think this one is overly verbose and probably falls out of context windows pretty quickly”
It likely will. Half way through a session I routinely watch the agent append my rules to the top of its thinking only to do exactly what it said it wasn’t going to do after another minute of thinking.
It will then apologize profusely right before doing it again.
As others have said, use hooks.
Love it! I think the power of LLMs to acquire new skills and deepen the knowdledge is underestimated.
When used correctly, they offer a huge advantage over those who don't use them and think they understand but remain superficial. I encourage you to ask even the most obvious questions.
This is not long at all compared to what is actually being used in production by my agents and others like Claude Code...
For those using Claude Code, I recommend Learning mode to instruct Claude to walk you through implementing the solution yourself rather than doing it for you. It’s very helpful when diving into a new domain, and helps build lower level intuition.
To enable it, run /config > output styles > Learning
Learning mode has been a huge help for me, it quickly became my favorite way to learn. I ended up created a “Coaching Mode” output style that took some of the learning concepts like stubbing todos for the users and added other intructions that better fit how I learn
This seems somewhat sensible to me - the genie _is_ out of the bottle, and students absolutely will use AI agents to finish assignments without learning a thing, but there is some value to showing how agents can be used as teaching tools and what healthy use _can_ look like
I think these are based on the one I posted a while back:
https://gist.github.com/1cg/a6c6f2276a1fe5ee172282580a44a7ac
Seems like a pretty close copy of Carson's (of HTMX fame) agent.md from 5 months ago
https://gist.github.com/1cg/a6c6f2276a1fe5ee172282580a44a7ac
This would be an interesting approach if the course supplied a custom Harness (perhaps in place of a textbook) and this was part of the instruction set inside of it. As a standalone thing you ask students to import into their agent, seems unlikely to work.
Hah, I like that these are presented as a CLAUDE.md.
(They have the same content duplicated in an AGENTS.md as well - I really wish Anthropic would hurry up and teach Claude Code to check for that file too.)
i think people out of school underestimate the power of exams. there's a huge difference in classes recently between ones with and without exams. if there is an exam, people are way more likely to study and therefore actually learn
This is such a realistic balance between completely banning coding agents and embracing the spirit of higher education
This is interesting. I don't know how the AI agent guidelines will be enforced because there will always be a model outside the curriculum that a student can use to bypass the guidelines. Encouraging academic integrity is useful but requires the student to buy into the idea that they are paying for an education, not a diploma. This is a tough problem and I have been wondering how CS departments are incorporating AI into the curriculum while encouraging appropriate use in a learning environment.
> What AI Agents SHOULD NOT Do
> * Run bash commands
Students who prefer to use zsh keep winning.
yeah I don't think that's going to work - it would be kind of like "we're releasing model answers to all assignments but please only use them as a teaching aid and don't copy from them"
best to
a) adapt assignments so that agents are bad at producing solutions
b) have more scenarios where students have to do things in controlled environments. Universities managed to adapt to 'any solution you need is readily available online' so I don't think it will be that different to have several times a month/year where students have to go into a room with nothing but pencil and paper to prove what knowledge they have vs what they have the skills to access
Calling an entity that's forbidden from acting on your behalf an "agent" seems funny but maybe it's meant as a catch-all term. Their use of "assistant" seems better for that purpose.
This is a very good baseline for future courses to build on, there would always be a group that wants to jailbreak this and thats okay, but have baseline agent support learning is needed in this ai first world.
I always thought it might be good to reframe educational questions in the age of LLMs. Something like "An apple falls on Sir Isaac Newton's head and a photon passes the event horizon of a black hole. Get from A to B." This way, it becomes physically impossible to just query for an answer because the student themselves will have to show the line of reasoning, ideally presenting to the class how they got from A to B.
It's a good idea, at the very least it communicates intent to students, but couldn't students just modify CLAUDE.md and not check in their changes to that?
When reading this I'm reminded of Charles Petzold's 2005 article "Does Visual Studio Rot The Mind?". https://www.charlespetzold.com/etcetera/DoesVisualStudioRotT...
And, yes students are going to follow it....
I am questioning the experience level of the TA if they think that AGENTS.md is going to work.
>Don't: Run bash commands
This seems unreasonable to me. One of the best uses of AI is that you can just tell your computer what to do in natural language and it does it. Running bash commands isn't part of the education, its busy work.
"Guidelines" eh?
Reminds me of this: https://www.youtube.com/watch?v=k9ojK9Q_ARE
I am really baffled by the comments in the spirit of "this is unenforceable, and therefore worthless".
I bet most people would not steal even if they knew they could get away with it.
Related:
CS336: Language Modeling from Scratch
Interesting. It makes me think of the idea of fighting piracy by providing a solid legal alternative through streaming platforms, etc.
I really like this. I'm currently doing a part time BSc and my current module explicitly allows AI usage as long as you 'cite it'. The guidelines are out of date in that they assume you are using a chatbot and not a coding harness. The temptation to have claude write all my pandas code has become too difficult for my self control, but at the same time I actively feel my education is suffering from using it. As I write my final paper I am thankful that I at least despise AI writing too much to use it for the actual marked assessment, but I still feel that I have cheated myself out of part of my education and probably wasted a lot of time going fast in the wrong direction because generating data frames, graphs, statistics, etc. is just so easy with claude
I just took a C1 Spanish class and it had almost exactly the same instructions. Hmmm and I do not wonder why...
I'm definitely going to use a variation of it for learning new programming languages.
Even though it seems radical, I think the right approach is to simply allow the students to use AI to its full potential, to generate answers, code, whatever.
The onus should be on the instructor to make sure that the student ends up actually understanding and being able to code/solve problems that they pose without using coding agents.
Why? Because:
1. this is exactly what is going on in the real world. People are able to get AI to do whatever the hell they want, but the ones who just use it lazily end up with huge cognitive debts and codebases riddled with opaque bugs that they do not understand whatsoever. If we prevent students from confronting this temptation, then we are sort of coddling or shielding them from it, and not really preparing them to avoid pitfalls of this type.
2. you can actually learn a LOT by being given the answer, if you actually care to learn. i personally think it's pretty fucking lame to handicap a student's ability to learn in an attempt to prevent lazy abuse. isn't the whole point of a grade to measure how well you understand things? can't you have pop quizzes, assignments on a computer with no agent use, written tests, etc etc. to catch the lazy abusers? this is an unnecessary prevention of lazy abuse that unfairly handicaps learning
What's the estimated RoI on doing this course?
Good practice for university of how to use agents in courses rather than forbidding those without distinction. Official AGENTS.md may be a new pattern in university courses.
Finally some sanity when it comes to AI.
I always wonder why there is such course. Using agent ai coding tool is trivial.
Is this all an elite educational institution with about $50bil in assets could muster, lol? This is completely and utterly unenforceable, and such, worthless.
There really needs to be diversity in delivery styles for different modules of courses according to their aims, with 'ai access' as a key variable.
If AI is allowed, it should be based on $x of usage/student, with an audit trail to prove no external funding was used, and module aims based on using AI to the max while conserving token use. Like actually creating wild, ambitious shit which takes cutting edge services to the max.
If AI is not allowed for a module, then it really needs to go back to the old skool, with handwritten exams, or coding using old machines and textbooks. Some skills, techniques, etc, really do need drilling.
Straddling the middle will help nobody, result in accusations, increase the burden on teaching staff, and result in a course without a realistic focus.
Though I guess if you're a big brand university, you don't really need to care about innovating. The money will keep pouring in. The whole further education sector is in dire need of a shake up.
this is like telling a student driver they have to push their car by hand instead of putting their foot on the pedal.
This is ridiculous. The genie is not going to go back into the bottle. This is the equivalent of "you wouldn't download a car". (Yes, we would.)
The solution is to scale the difficulty of the objective measures. Expect far more from students.
Reorient the university around physical laboratories and timesharing resources no single student could afford. It's already like this in many STEM disciplines.
More internships, more networking, more large projects. Less trivial tests of knowledge and credentialism.