So, to dodge this, I created a detailed prompt asking claude sonnet 4 to do this for me. I gave it very clear instructions (~2 min writeup) + the plaintext terraform plan. It successfully generated all the terraform state mv commands I needed in about 1-2 minutes, which I could xargs and boom, done! High value LLM query! Low amount of input tokens (relatively speaking), low amount of output tokens, saving magnitudes of time (which, as we all know, is money).
Please share your high-value LLM queries!
Any time that I’m trying to think through something or want an “opinion” about design choices or am I misting something, I type those two words in so it will be critical.
My next favorite prompt, “I’m having an issue with $X and having a hard time tracking it down. Help me work backwards. Don’t assume anything. Ask me clarifying questions as needed”. It’s great for rubber ducking.
For AWS troubleshooting, I ask if to give me AWS CLI commands to help it help me to debug and to always append “ | pbcopy” to it so I can just paste the output.
it one-shotted 600 lines of code which did the job perfectly. it understood from context the center of body, how to calculate the body normal, to rotate each point around that, all while handling edge cases to avoid errors. would've takens me hours if not days to tweak it manually to work.
"In 4 sentences, how would you do x".
"In 2 paragraphs summerise the pros and cons of y".
Not really specific coding tasks, but ask these types of questions a lot because often I'm not trying to be an expert or deeply understanding something but get a feel for the consensus view.
LLMs tend to be verbose by default.
In terms of coding I often ask, "Don't make changes, but how would you improve this piece of code?" Or "Don't make changes, but what's wrong with this test?".
I find Cursor at least loves to make changes when I didn't really want it to. I was just asking for some thoughts / opinions.