Check the finished-looking stuff hardest
Anthropic published a study of about 10,000 AI conversations and found something counterintuitive.
When people use AI to build things (code, apps, documents), they put more effort into directing it up front but less effort into evaluating what comes back. They’re more specific about what they want, and less likely to question the reasoning, check facts or catch what’s missing.
It makes sense. If the code runs or the document looks clean, your brain moves on. You don’t dig into whether the logic holds up. But this is backwards. The stuff you’re building and shipping is exactly the stuff that deserves the most scrutiny.
A few things that have changed how I work with AI:
- I treat every first output as a rough draft, even when it looks finished. Especially when it looks finished.
- I front-load instructions like “flag what you’re uncertain about” and “push back if my assumptions are wrong.” This completely changes the dynamic. Only 30% of users do this, according to the research.
When I’m building something, I make myself ask “what’s wrong with this?” before I ask “how do I ship this?”