another thing is that prompt optimisation increasingly feels like the wrong abstraction boundary. often the bottleneck is actually the tool schema or context assembly. optimising prompt text alone can leave most of the actual search space untouched
been revisiting prompt optimisation and increasingly think a lot of it is becoming specification compression. frontier models can already infer constraints we’d previously spell out, so adding instructions to patch individual eval failures can actually introduce interference
there’s also the question of whether shorter prompts generalise better or just overfit the eval set. removing redundant constraints can improve performance, but without held out evals it’s difficult to distinguish genuine compression from benchmark specific optimisation
was speaking to a mechanic today with 35 years of experience, trying to understand how they do things. at the end i went through everything i’d understood just to check i got it right and he goes “you’re hired”
made my day honestly
AGI - Allah guided intuition
ASI - Advanced Sharia intelligence
LLM - Long Live Muhammad
CoT - Chain of Tafsir
MCP - Muslim community protocol
GAN - Generative Arab Network
RAG - Relentless Allah Glorification
SDK - Sharia development kit