When product wanted to build a bot at our hackathon, I asked one question: How often does this come up? The answer changed their plans. Here's my five-question framework for deciding what's actually worth building and what decisions you need to make along the way.

AI acts more like a human than we admit. Give it imprecise instructions over a long conversation and it gets confused, same as a person would. Except a person who fixes an error, fixes that error. AI tends to rewrite the whole thing. So I built a skill that makes it prove the work is actually done, category by category, before I take “looks good” on faith.

I used to file client contracts by jurisdiction. Then I hit a client who started in the UK and didn’t stay there, and I stopped and thought: wait, where did I file this again? That hesitation was the first sign I’d built the whole structure wrong. Here’s what I built instead, and the one problem I haven’t fully solved yet.