A game changer for me in the last few weeks: asking Claude to prove completion through screenshots or video.
For smaller tasks, you can just inspect the code or run it locally. But for bigger projects, asking for proof — "show me a screenshot that this works" — actually prevents so many false positives.
I'd open a completed task and something wouldn't work. Or it would partially work. Or the AI would have made an assumption that seemed right in isolation but broke the broader system.
With screenshots, I can see instantly whether the result is plausible. Did the UI actually render correctly? Does the form actually submit? Did the data actually save?
This is simple but profound. It's the difference between "the code compiles" and "the feature works."
The AI can rationalize its way through code-level correctness all day. But when it has to prove it visually, when it has to show a working screen, the standard of evidence changes.
It's closer to how humans work. We trust what we see more than what we're told. Why shouldn't we apply that same principle to AI-generated work?
Fewer false positives. More confidence in deployment. Better integration of AI output into real systems.
Part of the #100DaysToOffload documenting agentic development in 2026