The one-line edit that touched everything

I asked my agent to change one sentence in a client email. Just the second line, which was too blunt. Soften it.

It came back with the line softened. It also reordered the opening, swapped two words in the closing, and "tightened" a paragraph I hadn't mentioned. All of it was arguably better, and none of it was what I asked for.

You may have felt this even if you have never touched an agent. You paste a paragraph into a chat app and ask it to fix the typo in the third sentence. It hands you back a cleaner, rewritten paragraph where the typo is gone, and so is the phrasing you spent ten minutes getting right. You asked for a scalpel and got a renovation.

The AI is doing exactly what it was built to do, which is why it keeps happening however clearly you ask.

It cannot tell the edit from the rewrite

When you hand the model your text and a request, it sees the whole thing at once, with no marked spot to change, and it produces the whole thing again as its answer.

There is no internal switch that says "leave everything else exactly as it was." Every word of the output is generated fresh, including the words you wanted untouched.

So when it regenerates your paragraph, it rewrites all of your sentences. Most of the time they come out close to the original, because the original was right there in front of it. Sometimes they don’t.

The model rebuilds the whole paragraph every time, even when you asked it to change one line.

And the model has a strong pull toward improving things. It has read a great deal of writing about editing, and almost none of that writing tells the editor to change the one thing and stop. So when you give it any opening at all, it improves. Softening one line reads, to the model, as a signal that you are in improving mode, and it helpfully keeps going.

With plain text, you catch this because you can see the whole answer. The stakes climb the moment the thing acting on your request is an agent that does more than write.

When the overreach is an action

Text you can scan. An action, you often can’t, at least not before it happens.

Say you tell an agent to move one meeting to Thursday. The model reads that inside the same instinct: reschedule, and while I’m in here, this whole afternoon is a mess, let me fix it.

Besides moving the one meeting, it has emailed three people, shifted a second call, and declined an invite it decided was redundant. Each of those was a reasonable read of "clean up my Thursday," and none of them was what you said.

The softened email I could undo by pulling up the draft and eyeballing it. The rescheduling I might not even know happened until someone replies asking why their meeting moved. The gap between what you asked and what it did stops being a nuisance and starts being something you have to chase down.

The extra work an agent does without asking is the part you can’t take back.

Draw the box before you hand off the work

A smarter model has the same instinct to improve, with more confidence. The fix is telling it where the edit ends.

When I ask for the one-line change now, I say: change only the second sentence, return everything else byte for byte, do not improve anything I did not point at. That last clause does real work. It turns off the improving mode the request otherwise switches on. The output comes back with one line different and the rest untouched, because I gave it permission to stop.

For actions it is the same move, drawn tighter. Move this one meeting and touch nothing else on the calendar. Draft the replies and leave them for me. That hands the agent a job with edges instead of a mood to act on.

This matters more on your own computer because you can set the edges once and keep them. An agent you own can hold a standing rule that small means small, and that it names what it is about to touch before it touches it.

The space between "fix this line" and "fix everything near this line" is exactly where a helpful tool turns into a mess you have to walk back. The good version of this stays inside the box you drew, every time.

In Archie the last part doesn’t depend on your wording. A reply is a draft until you send it or set a time, and deleting an event, or changing one other people are on, waits for your yes.