Goal, limits, finish line
The five parts
Goal: what you want and why. Context: the files, sites or accounts to use. Limits: what it must not do (spend money, send anything, touch certain folders). Finish line: how it knows it's done. Output: the format you want back, such as a table, a draft or a short report of what it did and anything it couldn't do.
Say when to stop and ask
Tell the agent which situations should make it pause: a login page, a payment screen, a choice between two plausible options, anything that looks like an error. "If unsure, stop and ask me" works better than hoping it guesses well. Agents tend to push on; a brief that permits stopping gets better results.
Make it report
Ask for a short log at the end: what it did, in order, what it skipped and why, and anything it was unsure about. This turns a black box into something you can check in a minute, and it shows you where to improve the brief next time.
Save and improve the brief
A brief that worked is worth keeping. Save it as a scheduled task, a skill, a Project instruction or just a note. Each time the agent goes wrong, add one line to fix that case. After a few runs you'll have instructions that reliably do the job, which is the real asset, more than any single result.
Write one sentence describing a task for an agent. Now list three things a literal-minded stranger could get wrong from that sentence.