Someone at work has quickly put together several internal tools using AI prompts. They seem to work, but I'm having trouble reconciling that speed with how long development usually takes. How should I think about the work and whether the tools are ready to rely on?
How do you get comfortable with AI-built tools at work?
Answers
Try separating how the code was produced from whether the tool is good. Defining the problem, checking the output, connecting the pieces and fixing mistakes are still work, even if a model wrote much of the code.
That doesn't mean a working demo is a finished product. The same questions apply as with any other internal tool: does it solve the right problem, and can the team understand and maintain it?
Reply to this answer
I'd focus the discussion on review rather than whether prompting counts as building. Before putting a tool into production, review its code and security, test the expected failure cases, and check it against your organization's policies and change controls.
Also ask who owns it, how it will be supported, and whether someone can diagnose and fix it when it breaks. A tool that works today can still be a headache if nobody understands what it does.
8c+brindlewrench_426:
That gives me a more useful conversation to have than arguing about whether the work counts.
Reply to this answer
Try it on a small script yourself. You'll get a better feel for both the speed and the places where the output needs correction. Keep it away from production until you've reviewed and tested it, though.
Reply to this answer
There's a separate question here from whether the tools are useful. Some people enjoy the craft of writing and understanding the code, and prompting can make the result feel less personally satisfying. It's reasonable to feel unsettled about that, or about what the change means for workers, without dismissing every tool it produces.
Reply to this answer
Add your answer
Answers are reviewed before publication.
q.518!fusemaraca:
Knowing what to ask for matters, but prompts can be ambiguous. Someone still needs to check the assumptions the model fills in.