As I browse around Reddit or Hackernews, I regularly see screenshots like this (link to post):
Or this (link to post):
I understand why people have vent at Claude/Cursor/whatever when it screws up. It talks like a person, so when it messes up, it’s tempting to tell it off, maybe even add a swear word or two.
However, I think not a helpful approach. Instead, when that happens, pause for a sec, and consider this:
The agent is going to forget the conversation once the session ends. Well, unless it has memory of some sort, in which case it might remember the wrong lesson.
What do you want the outcome of the conversation to be? Will yelling at the agent solve your problem? Will it prevent the mistake from happening again?
The agent is not a human. It’s a program that just demonstrated a bug. And you’re a software engineer. Why don’t you do some debugging?
So rather than typing “what the fuck did you just do??”, try some version of this prompt:
It looks like after the last prompt, {describe the issue}. Investigate why this happened and try to find the root cause. Suggest approaches to avoid it in the future.
Or just step back and think through how the issue could have happened—based on what you know about LLMs so far. Could it have been context rot from too many instructions? Or was important context not added somewhere? Should you have used a smarter model or set the reasoning level higher? Or maybe you need to update the rules in AGENTS.md/CLAUDE.md?
In the last year, I’ve now run twice into a situation where I was seriously tempted to type “WTF???” as my next prompt. However, after doing some investigating, the first issue led me to change the way I prompt and write skills, and the second issue turned out to be a bug.
Finally, keep in mind that LLMs are non-deterministic systems, and there’s an inherent risk that sometimes they’ll do something random and stupid. So it’s a good idea to put some boundaries in place to prevent them from causing real damage, such as blocking force-push on the repo’s main branch, or blocking write access to the prod infrastructure.