The line that lets it run without checking in every 2 minutes, the boundary that keeps it away from anything you can't undo, and 5 first jobs that are safe to hand over.
GPT-6 Astra doesn't just hand you text any more. It can open your apps, click, type, and stay on one job for a long stretch. Asked for a portrait in Canva, it opened the Canva editor and built it by hand over about an hour instead of calling an image generator.
The part nobody mentions: a model willing to work for an hour is also willing to stop nine times to check with you. You end up babysitting the thing you handed over.
This page is the fix for that, plus 5 first jobs that are safe to hand it and the checks to run before you connect it to anything.
Put this at the top of any long job. The first half tells it to act instead of checking in. The second half is the boundary that keeps it away from anything you can't take back.
Use the whole thing. The boundary is what makes the first half safe.
Infer my intent and task scope from my prompt and our prior conversation. Bias towards action, make reasonable assumptions, and work autonomously towards my goal. Only pause for confirmation when an action is clearly destructive or irreversible.
Without it you have told an AI with access to your computer to stop asking permission. With it, it still stops before anything it cannot undo: sending, paying, deleting, publishing, overwriting.
Never use the first sentence on its own.
Every one of these is something you could undo in a minute if it went sideways. That is the whole selection rule. Do one of these before you hand it anything that matters.
It is slow, so expect to wait. On a test of real computer tasks it averaged about 40 minutes a job.
Why it's safe: renaming is reversible and you can see the result instantly. Point it at a copy of a folder and give it the naming pattern you want.
Rename every file in [folder] to the pattern [YYYY-MM-DD-client-description]. Read the date from the file contents where you can. List anything you could not confidently rename instead of guessing.
Why it's safe: the tedious part is the typing, and you review before anything is submitted. Tell it explicitly to stop before the submit button.
Take each row in [spreadsheet] and fill in the matching fields on [the form]. Do not submit anything. Stop when the form is filled and tell me what you could not match.
Why it's safe: work on a duplicate. This is the job that used to eat an afternoon and produces something you can eyeball in seconds.
Work on a copy of [document]. Make the headings consistent, fix the list formatting, and match the fonts throughout. Do not change any wording. Tell me what you changed.
Why it's safe: it reads and writes to a new file, touching nothing that already exists. This is the one most solos feel immediately.
Read every receipt in [folder] and build a new spreadsheet with date, vendor, amount, and category. Flag any receipt where you could not read the amount rather than guessing it.
Why it's safe: sorting and copying, never deleting. Say so explicitly, because the fastest way to tidy is to delete and you do not want that.
Sort the images in [folder] into subfolders by [year and event]. Copy, never move or delete. Leave the originals exactly where they are.
Computer use means it acts inside apps you are already signed in to. Run these before the first real job, not after.
These come from the same round of Astra guidance and are aimed at people writing code or running agents. If that is not your week, ignore them.