When I first started building AI assistants in Appspace, my instinct was to load it up with every tool I could think of on Day 1. Search this, fetch that, summarise the other. The thinking was, give it everything and it'll be brilliant.
It wasn't. It got slower, the tool selection got fuzzier, and half the tools I'd carefully built barely got called.
The hack: ship your assistant with one tool. The most obvious, highest-value one. Nothing else.
Then build an AI Usage dashboard for your assistant, a week later and look at what people actually asked. That tells you tool #2. Watch it for another week, that tells you tool #3.
You end up with an assistant that's faster, sharper, and built around what your users genuinely need, not what you guessed they'd need on Day 1.
Bonus: tool descriptions matter more than you think. Short, action-led, specific. The model picks tools based on the description, not your system prompt.
#HindsightHack