Understand before you approve: what, why, is it safe
In the previous post I wrote about building the site in Antigravity: plan first, 12 stages, one commit per stage. This part is different. Now the subject is the click: what I read before I say yes.
What I wanted to solve
I work during the day, and at night and on Sundays I build with AI agents. I have never programmed professionally. A code editor with artificial intelligence asks for approval all the time. Install a program. Copy a file. Delete a folder. The button is fast. So is the damage.
During the construction of the site I made a rule for the Antigravity agent. Before each command that needs my approval, explain in one or two sentences three things: what the command does, why it is needed, and whether it is safe. Only then do I approve.
I also set this limit. No administrator password. Everything installed in the user folder.
I did not want to become the person who hits yes on autopilot.
Quick summary
- Before each command that asks for approval, the agent explains in one or two sentences: what it does, why it is needed, whether it is safe.
- No administrator password. The work stays in the user folder.
- Approving without understanding is the most common way for an agent to cause damage.
- Each approval becomes a short lesson. The rule is easy to copy.
First: what a command is (and what approval is)
A command, in this text, is a request the computer carries out. Install. Copy. Delete. Generate the site pages.
Approval is the moment the editor stops and waits for the human. In Antigravity that shows up on the screen. I operate it. The agent does not click for me.
Antigravity is Google’s code editor with artificial intelligence. I describe what I want. The model suggests files and commands. I read and decide.
Picture someone at the door with a box. Before coming in, they say what they brought, why they need to enter, and whether it touches only the box or the whole house. You listen, you look, and only then you open.
Inside, it is drier. The model builds the request. The editor shows the request and waits. If I say yes, the system runs it. The door analogy stops working at one point: the person at the door has a face. The command is a line of text. If you train the habit of approving without reading, the next line can be the wrong one, with the same confident look.
That is why the explanation has a ceiling: one or two sentences. Long text gets skipped when you are tired. A short sentence forces the agent to pick what matters. If the essential point does not fit, the request is still fuzzy. Then I do not approve.
First: administrator password and the user folder
Administrator, on a computer, is the account with permission to change the whole system. Install something in the core of the machine. Change a setting that applies to everyone who uses that computer. Delete things that are not in your folder.
An administrator password is the key to that account. When a program asks for that password, it is asking to leave your room and walk through the rest of the house.
The user folder is the corner of the computer that belongs to your account. Your documents, your projects, what the system treats as yours. A program installed there usually applies to you, not to the whole system.
Picture the desk drawer, not the building’s electrical panel. You touch what is yours.
Inside, the file system keeps your account’s files apart from the system’s files. The house analogy helps because the reach changes: the user folder is your space; administrator is the building. It fails at one point: even in your room, you can still throw away the only pad of paper. The password lock does not stop every mistake. It stops the mistake of opening a door a personal site did not need.
A paper drawer also does not run anything on its own. A command in the user folder still copies, deletes, and installs. That is why the three sentences stay necessary even without the password.
I asked for the work without that password. Everything in the user folder.
What the three questions catch
The first question is what. Without it, I approve an empty verb. “Update” can mean changing one letter or deleting a folder.
The second is why. Without it, the agent may be solving a problem I do not have. Needed for the request I made, not needed “because it looked cleaner”.
The third is whether it is safe. I do not take that at face value. “It is safe,” said by the agent itself, is its opinion, not a report. I want the risk in plain language: does it touch only this file? does it ask for an administrator password? does it delete anything? If the answer does not name the risk, I ask again.
A day-to-day example: if the request is to copy a project file, the explanation needs to say from where, to where, and that the original stays. If it does not say that, I refuse.
What I used
I used Google Antigravity, with me at the keyboard. I used the agent there, with the written rule: three questions, one or two sentences. I used the user folder, with no administrator password.
I am not pasting a command in this text. The habit is the point, not the line.
I use the tool. I am not selling the editor. I am telling you the rule I imposed during that construction.
How it went
If you want to copy the rule, this is the path:
- Write to the agent: before each command that needs approval, explain what it does, why it is needed, and whether it is safe, in one or two sentences.
- Read the three answers. If one of them is vague, refuse and ask again.
- Check that the request stays in the user folder. If it asks for an administrator password, the answer is no.
- Approve one command at a time. Only click when you can repeat, in one sentence, what will happen.
- If you refuse, say why. The next explanation needs to be clearer, not longer.
When it works, I can repeat what is about to run. Then I approve. If I cannot, I do not approve.
What I learned
I do not have a story about a deleted folder. I have the rule I wrote so that story would not happen.
The lesson for beginners: each approval becomes a short lesson. Copy the three-question rule for your agents.
If the agent answers the three questions with jargon, I do not approve. Without the risk named, I am left with an “it is safe” that does not help me.
The rule does not replace judgment. A command can arrive with a clear explanation and still be a bad idea. The three sentences give me material. The decision stays mine.
None of these tools replace judgment. They shorten the path between “I know what I want” and “it is live”.
What I would do differently: paste the rule into the agent at the start of the work, not in the middle. It was born during construction. It worked. Even so, a house rule is better if it exists before the rush.
Costs and limits
There is no money figure to report in this part. The rule does not cost extra. The cost is time: reading one or two sentences on each request. That is small next to undoing a wrong command.
What the rule does well: force a pause. Keep the install in the user folder. Turn each yes into a one-line lesson.
What it does not do: make a dangerous command safe just because it came with a paragraph. It also does not make the agent deserve an administrator password. In this way of working, that password does not come in.
This is what I did while building the site, not a study. The line “approving without understanding is the most common way for an agent to cause damage” is the reason for the choice. It is not published research.
Who it is for, and who it is not for
This helps if you use an AI editor that asks for approval and you are not a professional programmer. It also helps if you have already hit yes while tired.
It does not help if you want to send one paragraph and let the agent click everything alone. I refuse that setup.
It also does not help if you want the explanation as decoration. If you will not read it, the rule only delays the damage.
Closing
I wanted to stop approving in the dark. The rule fits in a note: what it does, why, whether it is safe. No administrator password. Everything in the user folder.
Do not approve what you cannot repeat in one sentence.
In the next post I will tell how the work was stored: site backup and records in Google Drive.
If you want to follow the journey, I am on X: @RodrigoesPorto.
Sources and links
- Google Antigravity: https://antigravity.google/
- This site: https://rodrigoesporto.com/