TL;DR:
Two questions sort most tasks: can you undo it, and how long would it take you to check the result?
Hand over reversible tasks that are quick to review. Keep irreversible ones, or approve each step yourself.
Agent features that do work in your apps are new. Treat them as early access and start small.
Suppose you could hand any chore on your computer to an assistant that clicks and types for you. Which would you give it first? Which would you never give it? Most people answer by gut feeling, and the gut is not a bad guide, but two plain questions make the answer much clearer.
The short answer
Ask first: is it reversible? If the assistant gets it wrong, can you undo it completely, such as by restoring a file, or does the mistake reach the world, such as a sent message or a payment? Ask second: what is the review effort? Once the work is done, how quickly can you check that it is right?
Reversible with a quick review is the best place to start. Irreversible with a slow review is the worst. The other two combinations sit in between and deserve a checkpoint where you look before the last step.
What happens step by step
List the task in one line, with what finished looks like.
Ask the reversibility question. Is there a copy, a trash folder or an undo? If not, it is irreversible.
Ask the review question. Could you check the result in under a minute, or would it take as long as doing it yourself?
Pick the setup. Hand over fully, hand over with a checkpoint before the final step, or keep it.
Start with a small version and widen the job only after the first one went well.
Walk-through: sorting ten tasks by risk
Here are ten ordinary tasks, sorted with those two questions. The verdicts are a judgment call, not a law, and your own situation may move a task up or down.
Task | Reversible? | Review effort | Verdict |
|---|---|---|---|
Summarise an article you are reading | Yes, nothing changes | Low: skim against the article | Hand over |
Rename 40 files in a folder, keeping a copy | Yes, copy exists | Low: scan the names | Hand over |
Rearrange columns in a copy of a spreadsheet | Yes, original is untouched | Medium: compare a few rows | Hand over |
Draft a reply to an email | Yes, it is only a draft | Low to medium: read it | Hand over the draft |
Send that reply | No, it is delivered | Low, but too late afterwards | Approve each time |
Fill in a form without submitting | Yes, until you press submit | Medium: read every field | Hand over, check before submit |
Submit the form | Usually no | Not applicable afterwards | Keep, or approve |
Delete old downloads | Only if the trash is kept | Medium: you must read the list | Checkpoint, review list first |
Pay a bill | No, money has moved | High: amount, payee, date | Keep |
Change a password or security setting | Hard to undo, can lock you out | High | Keep |
Count the verdicts: three are plain hand over, and two more are hand over with a limit (the draft and the unsubmitted form). That is five tasks you could start without much worry (3 + 2 = 5), and five that need you to stay in the loop (10 − 5 = 5): sending, submitting, deleting, paying and changing security settings.
Notice the pairs. Drafting a reply is safe, but sending it is not. Filling in a form is safe, but submitting it is not. A big job often splits into a safe half and a risky half, and the sensible move is to hand over the safe half and keep the button press.
What it can't do
Know what is risky for you. A file that looks like clutter may be the only copy of something important. You have to judge that.
Guarantee it does the task exactly as you meant. Agents misread screens and make wrong turns, which is why the review step matters.
Make an irreversible task reversible. Asking it to be careful does not unsend a message.
Replace a backup. If a task could damage files, have a copy before it starts.
Watch out: Never hand over your passwords or card numbers. Enter those yourself, every time.
Privacy and control
An agent working in your apps can see what those apps show. Before handing over a task, close windows that hold things unrelated to it. Prefer a setup where it must ask before it sends a message or pays for something, and where you can stop it at any moment.
DuckyHelper's agent features, which work in other apps with their own cursor, 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. Its everyday help of answering out loud and drawing on your screen does not change anything on your computer.
Try it yourself
Write down ten chores you did this week. Give each one the two questions and mark it hand over, checkpoint or keep. Choose the safest one and try it with an assistant, then check the result against your own version. If the check took less time than doing the task, you found a good fit.
Frequently asked questions
Is a quick review always enough?
Not for irreversible tasks. A quick review after a payment cannot take the payment back, so for those approve the step before it happens.
What is the safest first task?
Something reversible that you can check in under a minute, like summarising a page or renaming files in a copy of a folder.
Get started
