x
Because you asked it to improve the writing, and adding is the easiest thing that looks like improvement. Ask for better and you get more: another example, a smoother transition, a closing paragraph restating what you already said.
Your draft was 400 words. The improved version is 600 and says the same thing more comfortably. That isn't editing. It's upholstery.
The fix is to stop asking for improvement and start asking for defence:
"Go through this line by line. For each sentence, tell me what it does that no other sentence does. If it doesn't do something unique, cut it and show me the version without it. Do not add anything. Return the shorter version first, then the list of what you cut and why."
Four instructions, and each is load-bearing.
"Line by line" forces sequential attention. Ask for a general edit and you get a general answer — a paragraph of advice about tone. Ask sentence by sentence and it has to actually look.
"What does this do that no other sentence does" is the real test. Most flabby writing isn't badly written; it's redundant. The same point arrives three times in three registers. This question catches that, where "is this good?" never will.
"Do not add anything" is the instruction people leave out, and it's the one that changes the output most. Without it, the default pull toward helpfulness reasserts itself — it cuts two sentences and adds three.
"Show me what you cut and why" keeps you in charge. You'll disagree with maybe a fifth of the cuts, and the ones you rescue teach you something about what you actually value in your own voice.
It's converting a taste question into a structural one — the same move behind making it show its working.
"Is this good writing?" is a judgement call, and a model answering it will mirror your framing back with encouragement. "Does this sentence do unique work?" has an answer you can check. Structural questions get useful answers; evaluative ones get agreeable ones.
Here's the checkable part, and it's the most useful thing in this post: the assistant's default is agreement, and agreement is a worse editor than any human who mildly dislikes you. If you ask whether your draft is strong, you will be told it is. Not through dishonesty — because "this is a solid piece" is the highest-probability continuation of a request for feedback on a piece. You have to design the prompt to make agreement impossible, which is exactly what asking for defence does.
Anything you'll publish. Blog posts, newsletters, LinkedIn. The cut version is nearly always better, and shorter posts get read further.
Proposals and pitches. Where padding actively costs money, because the reader is looking for a reason to stop.
The about page. Notoriously the most over-written thing any business owns, because you're too close to it.
Not first drafts. Draft messy and long deliberately; run the editor afterwards. Editing while writing produces careful, boring prose. Get it out, then cut it down.
Once the cut version exists, one follow-up sharpens it further:
"Now find the three weakest sentences that survived and tell me what would have to be true for each to earn its place."
That surfaces the sentences you're attached to for reasons that aren't about the reader — the clever phrase, the qualifier protecting you from an objection nobody will raise, the throat-clearing before the actual point. It's the same instinct as making AI your most sceptical customer, pointed at your own prose instead of your offer.
Not a draft — something live. Run the prompt on your homepage copy or your last newsletter. The cut percentage on published work is usually humbling, and it's the fastest argument for making this a habit.
Save the prompt somewhere it doesn't have to be retyped — custom instructions is the natural home.
For the full prompt set, the free 5-min read is here: The Weekend Automation Plan.
Related Decodes
TL;DR: Asking AI to improve your writing makes it longer, because adding is the easiest thing that resembles improvement. Ask it to defend instead: go line by line, say what each sentence does that no other does, cut anything that fails, add nothing, and list what was cut and why. This works because it converts a taste question into a structural one — evaluative questions get agreeable answers, structural questions get useful ones.
