How to Find Tasks to Hand Off to AI

Whatever version of the prompt you use, you'll usually get back a short list of candidate tasks, not one obvious winner. Resist the urge to try all of them at once. Pick a single task, ideally one that's repetitive, has a clear right answer, and wouldn't be a disaster if an early attempt came out wrong. A weekly status summary is a better first try than anything touching a client contract or a paycheck.

Before you hand that task off to anything, write down what a good result actually looks like. Be specific: not 'a good summary' but 'includes the three numbers my manager always asks about, flags anything that changed since last week, and fits in one screen.' A standard written down in advance is the only thing that stops you from grading the AI's work against a feeling instead of a fact.

Then run it for two weeks, not one attempt. A single try tells you whether the AI can do the task once, under ideal conditions, with you paying close attention. Two weeks tells you whether it holds up on a day when the input is messier than usual, a step gets skipped, or something unusual comes through that the first version never saw.

Check every output against something other than the AI: the original source document, a person who already knows the right answer, or a simple test you can run yourself. An AI that reviews its own work is grading its own homework, and a second pass from the same system that made the mistake will often miss it again.

At the end of the two weeks, keep the handoff only if you'd be comfortable never undoing a single output it produced. That's a higher bar than 'it saved time,' and it's the right one: a task worth automating is one where a wrong answer getting through once wouldn't actually hurt anyone.

It's worth sitting with why this specific prompt caught on the way it did. Most advice about working with AI assumes you already know which task to hand off, and tells you how to do it well. This prompt flips that: it asks the AI to find the task for you, which is a genuinely useful thing to ask for, and also the exact place where an assistant with no real information about your job is most likely to guess.

That's not a reason to skip the plain version. It's a reason to treat its answer the way you'd treat a first guess from a new coworker who's trying to be helpful before they actually know how your job works: worth listening to, not worth acting on without checking. The follow-up, whether that's a memory-enabled conversation, a pasted week of real work, or a team's own task list, is what turns the guess into something closer to an actual finding.

The same caution applies to the list of candidate tasks you eventually get back, even from a well-informed version of the prompt. An AI pattern-matching across a description of your work can surface something that sounds exactly right and still be wrong about the details that would make it safe to automate. That's what the two-week test and the outside check in the previous sections are actually for.

Two weeks in, you should have a small, specific record: how often the output matched the standard you wrote down, what kind of input tripped it up, and whether checking the work took meaningfully less time than doing the task yourself. That record, not a general feeling that it's "going fine," is what should decide whether the handoff continues, gets adjusted, or goes back to being your job.

If it worked, resist expanding it to five more tasks right away. Let it run a little longer with the same level of checking, then widen it one task at a time, using the same before-you-hand-it-off standard each time. The habit of writing down what good looks like, testing for real, and checking against something other than the AI is the part that actually compounds, far more than any single task you hand off this month.