Your first task should teach you how the agent behaves without making a mistake expensive. A good starter task is not necessarily trivial. It is bounded.
Use the three-question test
Before delegating, ask:
- Can I name the area it should touch? “The homepage heading” is better than “the site.”
- Can I undo it cheaply? Text, styles, a small test, or documentation are good starting surfaces.
- Can I tell whether it worked? A build, a focused test, a rendered page, or a small diff gives you an observable finish line.
If one answer is no, make the task smaller before making the prompt longer.
Start low on the risk ladder
Good first tasks include explaining an unfamiliar folder, updating a contained piece of copy, adding a focused test, fixing a reproduced visual bug, or documenting a command that already works.
Save production credentials, irreversible data changes, broad dependency upgrades, and “redesign everything” work for later. Those tasks can be valid; they simply require more context, stronger recovery, and better verification.
This really happened. An agent working through a batch of downloads hit a login wall on the very last item, and the tempting move was to dig the credentials out of the browser to push through. The right move, and the rule that came out of it: when something is locked, skip it and report it. It is entry AF-10 on the watch-out list.
The best first session ends with you understanding the agent better, not with the agent touching the largest possible surface.