A prompt that works great in a demo and a prompt that survives thousands of unattended production runs are two very different documents.
The shift that mattered most for me: stop writing prompts as one giant paragraph of vibes and start writing them as structured contracts — clear role definition, explicit tool-use rules, worked examples for the edge cases that actually recur, and hard boundaries for what the agent should refuse or escalate instead of guessing. Skills and personas get modularised so they can be tested and versioned independently instead of copy-pasted across agents.
The biggest lesson: most 'the model is wrong' bugs are actually 'the prompt left this ambiguous' bugs. Once I started treating prompts like code — reviewed, versioned, and regression-tested against real transcripts — agent reliability went up more than any model upgrade did.