I’d always ask for a tldr, then just do some brief testing. Most of my prompts would run for an hour or more, so it was really just checking & responding to a few windows every hour while I do my other work.

It was always small tasks, given way too much compute, just so I could always be confident in the result. I always gave strict requirements, and if I was really lazy I’d say “iterate over this 10 times once finished, test everything, only come back to me with a blocker or if you have an extremely high level of certainty” some crap like that

It was a really lazy way to use AI, but it just worked well for me for the small tasks I needed to do, knew how, but wanted them automated, or if it was a once off - I just couldn’t be fucked, or busy with something else so I’d just ask a new terminal to do it.

Eg. My supplier is asking for sales forecast, fetch my sales from Shopify, look at seasonality from my stock-app and build a report and email to xxx@mysupplier.com - triple check it against Shopify, xero, report any anomalies

Then just blast that out on fable ultra code and it would rarely mess things like that up… I mean opus would hardly mess things like that up, but I’m just so short on time that I palm off any task I can, even tasks that might take just as long telling Claude but it’s much funner getting Claude to do it than repeating some process I’ve done a thousand times