Prompting is getting simpler as models get smarter

The Claude 5 context engineering thread is generating real debate about what good prompting actually looks like now. The emerging consensus: the elaborate, multi-page system prompts that defined 2023-era LLM work are becoming counterproductive. Builders are reporting that telling Claude to 'use your own judgment' on things like test generation works better than writing out every edge case. The thread frames this as a natural endpoint where a system prompt could eventually just say 'you know what to do.'

The pattern here is that prompting is becoming less like programming and more like delegation. You describe intent, not procedure. That is a meaningful shift for anyone building on top of these models, because it changes the skillset required and makes prompt libraries as a competitive moat much harder to sustain.

The counterpoint raised in the thread: this only works if you trust the model's defaults, which means the risk surface moves from 'did I write the right instructions' to 'do I actually know what the model will do without them.' For production systems, that is a scarier unknown.