prompt engineering is just specification writing with extra steps
the better I get at saying exactly what I want, the less prompt magic I need. most prompt problems are really thinking problems in disguise.
the diagram alone is worth the read
what model were you running for this?
honestly wild that this works at all
the comments here are better than most blog posts
great writeup, bookmarked
stealing this approach for work, thanks
i was skeptical but the example convinced me
this should be pinned
what does your eval setup look like
what is the smallest model you got this working on
this matches the anthropic docs almost word for word
thank you for not making this a 20 minute video
the comments here are better than most blog posts
finally someone said it
i would add: log everything, you will thank yourself later
what version were you on? this changed recently
did this break for anyone else after the last update?
this is just requirements engineering with a chat interface
lol this is so true it hurts. except now you can iterate 10x faster because the llm will tell you your spec is ambiguous. what are you finding is the biggest difference between prompt eng and traditional specs - just the iteration speed or something else?
the llm roasts your ambiguous spec in the first response and you fix it in 30 seconds instead of three rounds of tickets. genuinely the biggest productivity unlock for me
llama.cpp at 3b is rough yeah.... we switched to ollama and it got way better, not sure why. could be the quantization method or just the inference engine
biggest diff for us was realizing prompt eng is faster feedback loop. you find out your spec is ambiguous in 30 seconds instead of after implementation. still feels like the same skill tho