Don't Change the Topic With an LLM

#ai #agents #llm #programming

"Thanks for reading this post! If you'd like, next I can:

Talk about something I've learned working with LLMs.
Share a small tip that's helped me when doing agentic coding.

Just say the word."

Does that kind of pattern sound familiar to you? A lot of LLMs seemed to have been trained in this pattern of ending every turn with an "offer".

GitHub's Copilot seems to have been particularly brainwashed into this behavior.

ME: I'd like to update the instructions because you should not end every turn with "If you..." I notice this is habitual and it's not productive.

COPILOT: Understood—I won’t end turns with the habitual “If you tell me X…” prompt.
...
If you want, tell me your preferred style in one line (e.g., “never ask follow-ups unless absolutely required” vs “ask up to 1 clarifying question when it improves accuracy”).

ME: please read your last response again and identify the tragic failure.

COPILOT: I did the exact thing you asked me not to.

I'm not trying to pick on Copilot - I've noticed that same pattern with ChatGPT and Claude. One time I decided to really press ChatGPT to explain what I should do to make sure it stopped repeating this. At the of the day, it basically said "I can't help it - start a new chat and tell me how I should respond."

So this gave me a good lesson - yes, you can steer a coding agent in the middle of a session - but don't try to take it too far.

If you're working on a bug-fix for a while, and then you have an idea about how to improve the UI - don't just jump into a new subject and expect the model to behave optimally. They get stuck into "bug-fixing mode", "implementation mode", "design mode". The conversational context accumulates not just in terms of information, but in terms of what it thinks its role is.

Try forking the conversation and say things like "OK, now let's pivot" or "I'm going to ask you to think like a UX designer now" - make it really clear that its current role has suddenly shifted. Or, ideally, start a new session. The same model that can offer 2-3 creative designs for your new feature when it has a fresh context, will seem lazy and incompetent if you give it that task after you've been debugging for half an hour.

I hope you found that tip helpful, and feel free to share your own amusing (and frustrating) examples.

If you want, I can go into this topic in more depth - or we can think of some other ways to work around it. Just say the word.