TL;DR:
A task an agent can finish has a goal, a place to work, a finish line and a few limits.
Vague tasks fail because the agent has to guess, and guesses are where mistakes come from.
Put the approval points in the task: where it must stop and ask.
"Sort out my inbox." It is how most of us would say it to a friend, and a friend would ask three questions before starting. An agent that cannot ask, or does not, will guess instead. Some of its guesses will be right. The ones that are wrong might archive something important. A clearer task takes a minute to write and saves much more.
Here is a simple structure for tasks an agent can finish, shown on a before-and-after.
The four parts
Goal. What should be true when it is done? Describe the result, not just the activity.
Where. Which app, folder, site or sheet should it work in? Name it.
Done when. A check you can apply: a count, a visible change, a file in a place.
Limits. What it must not touch, and what it must ask about first.
Walk-through: rewriting a vague task into a clear one
The vague version: "Sort out my inbox."
Questions an agent cannot answer from that: Sort by what? Which inbox? Delete things or only move them? What counts as done? What about messages that need a reply?
Now rewrite it with the four parts:
Goal: "Move newsletters out of my inbox so I can see real messages."
Where: "In my work email inbox, only messages from the last 30 days."
Done when: "The inbox has no newsletters from the last 30 days, and I get a list of what you moved."
Limits: "Move, never delete. Do not reply to or send anything. If you are not sure a message is a newsletter, leave it and list it for me."
The second version is five lines, takes about a minute to write, and an agent can follow it without guessing. If something goes wrong, you can see exactly which part was not met.
Why each part matters
Goal stops it from doing a different job well.
Where stops it from working in the wrong place. A personal inbox and a work inbox look alike.
Done when stops it from finishing too soon or never finishing. It also gives you a way to check.
Limits stop the damage from the worst case. "Move, never delete" turns a possible disaster into a tidy-up.
Check that it is finishable
Test your task by asking three questions. Could someone who has never met you do it from this description? Could you tell, by looking, whether it was done? Is there any step where a wrong guess would be costly? If the third answer is yes, add an approval point there.
Tip: Put the stop points in the task. "Show me the list before you move anything" costs little and catches the biggest mistakes.
Give an example when you can
If a task depends on judgment, show one case. "Newsletters are things like the weekly digest from the school and the shop discount emails" gives the agent something concrete to match against. One or two examples often do more than a paragraph of rules, and they make it easy to say what you meant when something is misjudged.
Equally, name a case it should leave alone: "emails from my teacher stay in the inbox, even if they look like bulk mail". Edge cases you mention up front are edge cases you will not have to repair later.
Ambiguity words to watch
Words like "clean up", "organise", "fix", "handle" and "deal with" hide a lot of choices. Replace them with the specific actions you want: move, rename, sort, list, copy. The more specific the verb, the fewer places for a wrong guess.
You can also ask the agent to restate the task back to you in its own words before it begins. If its version leaves out a limit or changes the goal, you have caught the misunderstanding for free. It is a thirty-second step that works on people too.
Over time, keep the tasks that worked well as templates. A good task with the four parts filled in can be reused with new names and dates, which saves effort and keeps your standards consistent from one run to the next.
What this does not guarantee
A clear task helps a lot, but pages change and software misreads things. Look at the result and the record of steps. If the result is off, adjust the task and run it again on a small batch.
Privacy and control
Name only the apps and folders the agent needs. DuckyHelper's agent features are in early access. It's built to ask first before it sends a message or pays for anything, and not to type your passwords or card numbers.
Try it yourself
Take a chore you would normally describe in five words. Write it out with goal, where, done when and limits. Then ask a friend to do it from your note. Wherever they hesitate, your task has a gap.
Frequently asked questions
Do I have to write this much every time?
No. Short tasks need less. Use the full structure for anything that changes things or is hard to undo.
What if I do not know what done looks like?
Start with a smaller task, such as listing what it would do, and look at that first.
Can the agent ask me questions?
A good one should, when something is unclear. Still, better instructions mean fewer interruptions.
Get started
